{ "pdf_info": [ { "para_blocks": [ { "bbox": [ 111, 103, 483, 121 ], "type": "title", "angle": 0, "lines": [ { "bbox": [ 111, 103, 483, 121 ], "spans": [ { "bbox": [ 111, 103, 483, 121 ], "type": "text", "content": "A skeletonization algorithm for gradient-based optimization" } ] } ], "index": 2 }, { "bbox": [ 69, 142, 525, 186 ], "type": "text", "angle": 0, "lines": [ { "bbox": [ 69, 142, 525, 186 ], "spans": [ { "bbox": [ 69, 142, 525, 186 ], "type": "text", "content": "Martin J. Menten" }, { "bbox": [ 69, 142, 525, 186 ], "type": "inline_equation", "content": "^{1,2}" }, { "bbox": [ 69, 142, 525, 186 ], "type": "text", "content": " Johannes C. Paetzold" }, { "bbox": [ 69, 142, 525, 186 ], "type": "inline_equation", "content": "^{1,2}" }, { "bbox": [ 69, 142, 525, 186 ], "type": "text", "content": " Veronika A. Zimmer" }, { "bbox": [ 69, 142, 525, 186 ], "type": "inline_equation", "content": "^{1}" }, { "bbox": [ 69, 142, 525, 186 ], "type": "text", "content": " Suprosanna Shit" }, { "bbox": [ 69, 142, 525, 186 ], "type": "inline_equation", "content": "^{1}" }, { "bbox": [ 69, 142, 525, 186 ], "type": "text", "content": " Ivan Ezhov" }, { "bbox": [ 69, 142, 525, 186 ], "type": "inline_equation", "content": "^{1}" }, { "bbox": [ 69, 142, 525, 186 ], "type": "text", "content": " Robbie Holland" }, { "bbox": [ 69, 142, 525, 186 ], "type": "inline_equation", "content": "^{2}" }, { "bbox": [ 69, 142, 525, 186 ], "type": "text", "content": " Monika Probst" }, { "bbox": [ 69, 142, 525, 186 ], "type": "inline_equation", "content": "^{1}" }, { "bbox": [ 69, 142, 525, 186 ], "type": "text", "content": " Julia A. Schnabel" }, { "bbox": [ 69, 142, 525, 186 ], "type": "inline_equation", "content": "^{1}" }, { "bbox": [ 69, 142, 525, 186 ], "type": "text", "content": " Daniel Rueckert" }, { "bbox": [ 69, 142, 525, 186 ], "type": "inline_equation", "content": "^{1,2}" }, { "bbox": [ 69, 142, 525, 186 ], "type": "text", "content": " Technical University of Munich Imperial College London" } ] } ], "index": 3 }, { "bbox": [ 143, 213, 192, 225 ], "type": "title", "angle": 0, "lines": [ { "bbox": [ 143, 213, 192, 225 ], "spans": [ { "bbox": [ 143, 213, 192, 225 ], "type": "text", "content": "Abstract" } ] } ], "index": 4 }, { "bbox": [ 47, 238, 290, 550 ], "type": "text", "angle": 0, "lines": [ { "bbox": [ 47, 238, 290, 550 ], "spans": [ { "bbox": [ 47, 238, 290, 550 ], "type": "text", "content": "The skeleton of a digital image is a compact representation of its topology, geometry, and scale. It has utility in many computer vision applications, such as image description, segmentation, and registration. However, skeletonization has only seen limited use in contemporary deep learning solutions. Most existing skeletonization algorithms are not differentiable, making it impossible to integrate them with gradient-based optimization. Compatible algorithms based on morphological operations and neural networks have been proposed, but their results often deviate from the geometry and topology of the true medial axis. This work introduces the first three-dimensional skeletonization algorithm that is both compatible with gradient-based optimization and preserves an object's topology. Our method is exclusively based on matrix additions and multiplications, convolutional operations, basic non-linear functions, and sampling from a uniform probability distribution, allowing it to be easily implemented in any major deep learning library. In benchmarking experiments, we prove the advantages of our skeletonization algorithm compared to non-differentiable, morphological, and neural-network-based baselines. Finally, we demonstrate the utility of our algorithm by integrating it with two medical image processing applications that use gradient-based optimization: deep-learning-based blood vessel segmentation, and multimodal registration of the mandible in computed tomography and magnetic resonance images." } ] } ], "index": 5 }, { "bbox": [ 47, 573, 128, 586 ], "type": "title", "angle": 0, "lines": [ { "bbox": [ 47, 573, 128, 586 ], "spans": [ { "bbox": [ 47, 573, 128, 586 ], "type": "text", "content": "1. Introduction" } ] } ], "index": 6 }, { "bbox": [ 46, 594, 289, 715 ], "type": "text", "angle": 0, "lines": [ { "bbox": [ 46, 594, 289, 715 ], "spans": [ { "bbox": [ 46, 594, 289, 715 ], "type": "text", "content": "Skeletonization algorithms aim at extracting the medial axis of an object, which is defined as the set of points that have more than one closest point on the object's boundary [7]. This lower-dimensional representation compactly encodes various geometric, topological, and scale features. As such, it is useful for many tasks in computer vision, including object description, compression, recognition, tracking, registration, and segmentation [18, 22, 25, 40, 45]. While there are several efficient approaches to calculate the medial axis in continuous space, extracting the skeleton of a discrete digital" } ] } ], "index": 7 }, { "bbox": [ 304, 213, 548, 299 ], "type": "text", "angle": 0, "lines": [ { "bbox": [ 304, 213, 548, 299 ], "spans": [ { "bbox": [ 304, 213, 548, 299 ], "type": "text", "content": "image is not trivial. The search for accurate skeletonization algorithms, which can process two-dimensional and three-dimensional digital images, has spawned a plethora of research works [5, 6, 8, 15, 18, 31, 42, 46]. For a comprehensive overview and taxonomy of skeletonization algorithms and their applications, we refer to the excellent survey by Saha et al. [32]." } ] } ], "index": 8 }, { "bbox": [ 304, 302, 548, 566 ], "type": "text", "angle": 0, "lines": [ { "bbox": [ 304, 302, 548, 566 ], "spans": [ { "bbox": [ 304, 302, 548, 566 ], "type": "text", "content": "Today, computer vision tasks are commonly solved using deep learning. Skeletonization may be used as building block or inductive bias in these image processing applications [13, 16, 38]. However, most established skeletonization methods are not compatible with backpropagation and gradient-based optimization [32]. The few works that have integrated skeletonization with deep learning pipelines rely on morphological skeletonization algorithms [38]. This class of algorithms is based on simple morphological operations, such as erosion and dilation [20, 44]. However, they will often result in breaks in the skeleton, causing it to diverge from the geometry and topology of the true medial axis. Recently, learning-based methods have also been harnessed for skeletonization [9, 13, 16, 23, 24, 26, 36, 37]. Most of these works train an encoder-decoder neural network on pairs of input images and ground truth skeletons, which have previously been obtained using a classical skeletonization algorithm. While learned approaches are intrinsically compatible with backpropagation, they are not guaranteed to preserve the topology of the input. Additionally, they are susceptible to domain shifts between the training and inference data [10, 43]." } ] } ], "index": 9 }, { "bbox": [ 304, 570, 550, 714 ], "type": "text", "angle": 0, "lines": [ { "bbox": [ 304, 570, 550, 714 ], "spans": [ { "bbox": [ 304, 570, 550, 714 ], "type": "text", "content": "Our contribution This work bridges the gap between traditional skeletonization principles with strict topology guarantees and their integration with gradient-based optimization. We introduce a skeletonization algorithm that is topology-preserving, domain-agnostic, and compatible with backpropagation (see Figure 1). Our algorithm is exclusively based on matrix additions and multiplications, convolutional operations, basic non-linear functions, and sampling from a uniform probability distribution, allowing it to be easily implemented in any major deep learning library, such as PyTorch or Tensorflow [2, 27]. In benchmarking experiments, we establish that our algorithm outperforms non-differentiable," } ] } ], "index": 10 } ], "discarded_blocks": [ { "bbox": [ 65, 2, 111, 34 ], "type": "header", "angle": 0, "lines": [ { "bbox": [ 65, 2, 111, 34 ], "spans": [ { "bbox": [ 65, 2, 111, 34 ], "type": "text", "content": "CVF" } ] } ], "index": 0 }, { "bbox": [ 146, 0, 494, 37 ], "type": "header", "angle": 0, "lines": [ { "bbox": [ 146, 0, 494, 37 ], "spans": [ { "bbox": [ 146, 0, 494, 37 ], "type": "text", "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." } ] } ], "index": 1 }, { "bbox": [ 293, 748, 318, 758 ], "type": "page_number", "angle": 0, "lines": [ { "bbox": [ 293, 748, 318, 758 ], "spans": [ { "bbox": [ 293, 748, 318, 758 ], "type": "text", "content": "21394" } ] } ], "index": 11 } ], "page_size": [ 612, 792 ], "page_idx": 0 }, { "para_blocks": [ { "type": "image", "bbox": [ 82, 71, 516, 233 ], "blocks": [ { "bbox": [ 82, 71, 516, 233 ], "lines": [ { "bbox": [ 82, 71, 516, 233 ], "spans": [ { "bbox": [ 82, 71, 516, 233 ], "type": "image", "image_path": "c4641c3c4532c86f7c90dccd712b7d316428289335b65eccff0692b0c8a4f04b.jpg" } ] } ], "index": 0, "angle": 0, "type": "image_body" }, { "bbox": [ 45, 238, 547, 283 ], "lines": [ { "bbox": [ 45, 238, 547, 283 ], "spans": [ { "bbox": [ 45, 238, 547, 283 ], "type": "text", "content": "Figure 1. Most existing skeletonization algorithms are not differentiable, making it impossible to integrate them with gradient-based optimization. Morphological and neural-network-based solutions can be used with backpropagation, but alter the topology of the object by introducing breaks along the skeleton. Our proposed skeletonization algorithm preserves the topology while simultaneously being compatible with gradient-based optimization." } ] } ], "index": 1, "angle": 0, "type": "image_caption" } ], "index": 0 }, { "bbox": [ 46, 302, 290, 387 ], "type": "text", "angle": 0, "lines": [ { "bbox": [ 46, 302, 290, 387 ], "spans": [ { "bbox": [ 46, 302, 290, 387 ], "type": "text", "content": "morphological, and neural-network-based baselines. Finally, we directly integrate our skeletonization algorithm with two medical image processing applications that rely on gradient-based optimization. We show that it enhances both deep-learning-based blood vessel segmentation, and multimodal registration of the mandible in computed tomography (CT) and magnetic resonance (MR) images." } ] } ], "index": 2 }, { "bbox": [ 47, 396, 130, 410 ], "type": "title", "angle": 0, "lines": [ { "bbox": [ 47, 396, 130, 410 ], "spans": [ { "bbox": [ 47, 396, 130, 410 ], "type": "text", "content": "2. Prerequisites" } ] } ], "index": 3 }, { "bbox": [ 46, 416, 288, 489 ], "type": "text", "angle": 0, "lines": [ { "bbox": [ 46, 416, 288, 489 ], "spans": [ { "bbox": [ 46, 416, 288, 489 ], "type": "text", "content": "Digital images A discrete three-dimensional image is an array of points " }, { "bbox": [ 46, 416, 288, 489 ], "type": "inline_equation", "content": "P = \\{p(x,y,z)\\}" }, { "bbox": [ 46, 416, 288, 489 ], "type": "text", "content": ", which are each assigned an intensity value, on a lattice defined by Cartesian coordinates " }, { "bbox": [ 46, 416, 288, 489 ], "type": "inline_equation", "content": "x,y,z\\in \\mathbb{Z}" }, { "bbox": [ 46, 416, 288, 489 ], "type": "text", "content": " [14]. Owing to the grid nature of the image, we can define the 6-neighborhood " }, { "bbox": [ 46, 416, 288, 489 ], "type": "inline_equation", "content": "N_{6}(p)" }, { "bbox": [ 46, 416, 288, 489 ], "type": "text", "content": ", 18-neighborhood " }, { "bbox": [ 46, 416, 288, 489 ], "type": "inline_equation", "content": "N_{18}(p)" }, { "bbox": [ 46, 416, 288, 489 ], "type": "text", "content": ", and 26-neighborhood " }, { "bbox": [ 46, 416, 288, 489 ], "type": "inline_equation", "content": "N_{26}(p)" }, { "bbox": [ 46, 416, 288, 489 ], "type": "text", "content": " of a point [30]:" } ] } ], "index": 4 }, { "bbox": [ 50, 495, 249, 509 ], "type": "interline_equation", "angle": 0, "lines": [ { "bbox": [ 50, 495, 249, 509 ], "spans": [ { "bbox": [ 50, 495, 249, 509 ], "type": "interline_equation", "content": "N _ {6} (p) = \\left\\{p ^ {\\prime}; \\left(\\left| x - x ^ {\\prime} \\right| + \\left| y - y ^ {\\prime} \\right| + \\left| z - z ^ {\\prime} \\right|\\right) \\leq 1 \\right\\},", "image_path": "a0e6e0aa477ad0939b247a619620afbdd0ef05ce1fa878bd00d18c604f12f708.jpg" } ] } ], "index": 5 }, { "bbox": [ 49, 510, 287, 524 ], "type": "interline_equation", "angle": 0, "lines": [ { "bbox": [ 49, 510, 287, 524 ], "spans": [ { "bbox": [ 49, 510, 287, 524 ], "type": "interline_equation", "content": "N _ {2 6} (p) = \\left\\{p ^ {\\prime}; \\max \\left(\\left| x - x ^ {\\prime} \\right|, \\left| y - y ^ {\\prime} \\right|, \\left| z - z ^ {\\prime} \\right|\\right) \\leq 1 \\right\\}, \\tag {1}", "image_path": "e566fc92ebbb09bf3febeb089fb76cbd5f6d90cc62cdc76ff2510939fe030ae6.jpg" } ] } ], "index": 6 }, { "bbox": [ 49, 525, 287, 539 ], "type": "interline_equation", "angle": 0, "lines": [ { "bbox": [ 49, 525, 287, 539 ], "spans": [ { "bbox": [ 49, 525, 287, 539 ], "type": "interline_equation", "content": "N _ {1 8} (p) = \\left\\{p ^ {\\prime}; \\left(\\left| x - x ^ {\\prime} \\right| + \\left| y - y ^ {\\prime} \\right| + \\left| z - z ^ {\\prime} \\right|\\right) \\leq 2 \\right\\} \\cap N _ {2 6} (p).", "image_path": "10395f5a676ef8f2254940ec71524f6f69a53bc82487a8254e2dbdc6b10a4fff.jpg" } ] } ], "index": 7 }, { "bbox": [ 46, 545, 287, 594 ], "type": "text", "angle": 0, "lines": [ { "bbox": [ 46, 545, 287, 594 ], "spans": [ { "bbox": [ 46, 545, 287, 594 ], "type": "text", "content": "Points that are inside each other's " }, { "bbox": [ 46, 545, 287, 594 ], "type": "inline_equation", "content": "n" }, { "bbox": [ 46, 545, 287, 594 ], "type": "text", "content": "-neighborhood are called " }, { "bbox": [ 46, 545, 287, 594 ], "type": "inline_equation", "content": "n" }, { "bbox": [ 46, 545, 287, 594 ], "type": "text", "content": "-adjacent. Two points " }, { "bbox": [ 46, 545, 287, 594 ], "type": "inline_equation", "content": "p" }, { "bbox": [ 46, 545, 287, 594 ], "type": "text", "content": " and " }, { "bbox": [ 46, 545, 287, 594 ], "type": "inline_equation", "content": "p'" }, { "bbox": [ 46, 545, 287, 594 ], "type": "text", "content": " are said to be " }, { "bbox": [ 46, 545, 287, 594 ], "type": "inline_equation", "content": "n" }, { "bbox": [ 46, 545, 287, 594 ], "type": "text", "content": "-connected, if there is a sequence of points " }, { "bbox": [ 46, 545, 287, 594 ], "type": "inline_equation", "content": "p = p_0, \\dots, p_k = p'" }, { "bbox": [ 46, 545, 287, 594 ], "type": "text", "content": " so that each " }, { "bbox": [ 46, 545, 287, 594 ], "type": "inline_equation", "content": "p_i" }, { "bbox": [ 46, 545, 287, 594 ], "type": "text", "content": " is " }, { "bbox": [ 46, 545, 287, 594 ], "type": "inline_equation", "content": "n" }, { "bbox": [ 46, 545, 287, 594 ], "type": "text", "content": "-adjacent to " }, { "bbox": [ 46, 545, 287, 594 ], "type": "inline_equation", "content": "p_{i-1}" }, { "bbox": [ 46, 545, 287, 594 ], "type": "text", "content": " for " }, { "bbox": [ 46, 545, 287, 594 ], "type": "inline_equation", "content": "1 \\leq i \\leq k" }, { "bbox": [ 46, 545, 287, 594 ], "type": "text", "content": "." } ] } ], "index": 8 }, { "bbox": [ 46, 594, 289, 714 ], "type": "text", "angle": 0, "lines": [ { "bbox": [ 46, 594, 289, 714 ], "spans": [ { "bbox": [ 46, 594, 289, 714 ], "type": "text", "content": "Euler characteristic In the special case of a binary image the value of each point is either 1 or 0. The foreground of a binary image is represented by the set " }, { "bbox": [ 46, 594, 289, 714 ], "type": "inline_equation", "content": "S" }, { "bbox": [ 46, 594, 289, 714 ], "type": "text", "content": " of points with value 1, while the set " }, { "bbox": [ 46, 594, 289, 714 ], "type": "inline_equation", "content": "\\overline{S}" }, { "bbox": [ 46, 594, 289, 714 ], "type": "text", "content": " denotes the remaining points with value 0. Based on above's definition of connectedness, we can define an object " }, { "bbox": [ 46, 594, 289, 714 ], "type": "inline_equation", "content": "O" }, { "bbox": [ 46, 594, 289, 714 ], "type": "text", "content": " as a set of " }, { "bbox": [ 46, 594, 289, 714 ], "type": "inline_equation", "content": "n" }, { "bbox": [ 46, 594, 289, 714 ], "type": "text", "content": "-connected points in " }, { "bbox": [ 46, 594, 289, 714 ], "type": "inline_equation", "content": "S" }, { "bbox": [ 46, 594, 289, 714 ], "type": "text", "content": ". An object in " }, { "bbox": [ 46, 594, 289, 714 ], "type": "inline_equation", "content": "\\overline{S}" }, { "bbox": [ 46, 594, 289, 714 ], "type": "text", "content": " that is completely surrounded by points in " }, { "bbox": [ 46, 594, 289, 714 ], "type": "inline_equation", "content": "S" }, { "bbox": [ 46, 594, 289, 714 ], "type": "text", "content": " is called a cavity " }, { "bbox": [ 46, 594, 289, 714 ], "type": "inline_equation", "content": "C" }, { "bbox": [ 46, 594, 289, 714 ], "type": "text", "content": " in " }, { "bbox": [ 46, 594, 289, 714 ], "type": "inline_equation", "content": "S" }, { "bbox": [ 46, 594, 289, 714 ], "type": "text", "content": ". Finally, a hole " }, { "bbox": [ 46, 594, 289, 714 ], "type": "inline_equation", "content": "H" }, { "bbox": [ 46, 594, 289, 714 ], "type": "text", "content": " can be intuitively described as a tunnel through " }, { "bbox": [ 46, 594, 289, 714 ], "type": "inline_equation", "content": "S" }, { "bbox": [ 46, 594, 289, 714 ], "type": "text", "content": ". Objects, cavities and holes can be analogously defined for " }, { "bbox": [ 46, 594, 289, 714 ], "type": "inline_equation", "content": "\\overline{S}" }, { "bbox": [ 46, 594, 289, 714 ], "type": "text", "content": ". By combining the number of" } ] } ], "index": 9 }, { "bbox": [ 305, 302, 547, 327 ], "type": "text", "angle": 0, "lines": [ { "bbox": [ 305, 302, 547, 327 ], "spans": [ { "bbox": [ 305, 302, 547, 327 ], "type": "text", "content": "objects, cavities, and holes the Euler characteristic, or genus, of a 6-connected set of points " }, { "bbox": [ 305, 302, 547, 327 ], "type": "inline_equation", "content": "G_{6}" }, { "bbox": [ 305, 302, 547, 327 ], "type": "text", "content": " can be determined [14]:" } ] } ], "index": 10 }, { "bbox": [ 376, 333, 545, 345 ], "type": "interline_equation", "angle": 0, "lines": [ { "bbox": [ 376, 333, 545, 345 ], "spans": [ { "bbox": [ 376, 333, 545, 345 ], "type": "interline_equation", "content": "G _ {6} = \\# O - \\# H + \\# C. \\tag {2}", "image_path": "e5e7c77082e9a85ec832a90875625ccfeb720ae5c799f30652de7b97753449de.jpg" } ] } ], "index": 11 }, { "bbox": [ 304, 352, 547, 412 ], "type": "text", "angle": 0, "lines": [ { "bbox": [ 304, 352, 547, 412 ], "spans": [ { "bbox": [ 304, 352, 547, 412 ], "type": "text", "content": "It is also possible to span a graph between all 6-neighbors. On this graph, simplicial complexes consisting of one, two, four, or eight points are called vertex " }, { "bbox": [ 304, 352, 547, 412 ], "type": "inline_equation", "content": "v" }, { "bbox": [ 304, 352, 547, 412 ], "type": "text", "content": ", edge " }, { "bbox": [ 304, 352, 547, 412 ], "type": "inline_equation", "content": "e" }, { "bbox": [ 304, 352, 547, 412 ], "type": "text", "content": ", face " }, { "bbox": [ 304, 352, 547, 412 ], "type": "inline_equation", "content": "f" }, { "bbox": [ 304, 352, 547, 412 ], "type": "text", "content": ", or octant " }, { "bbox": [ 304, 352, 547, 412 ], "type": "inline_equation", "content": "oct" }, { "bbox": [ 304, 352, 547, 412 ], "type": "text", "content": ", respectively [14]. " }, { "bbox": [ 304, 352, 547, 412 ], "type": "inline_equation", "content": "G_{6}" }, { "bbox": [ 304, 352, 547, 412 ], "type": "text", "content": " can also be calculated based on the number of these complexes via" } ] } ], "index": 12 }, { "bbox": [ 366, 418, 545, 431 ], "type": "interline_equation", "angle": 0, "lines": [ { "bbox": [ 366, 418, 545, 431 ], "spans": [ { "bbox": [ 366, 418, 545, 431 ], "type": "interline_equation", "content": "G _ {6} = \\# v - \\# e + \\# f - \\# o c t. \\tag {3}", "image_path": "a6dda8e6bc9eee4a51bb4741df73708397ecaea8bdae12bb6a1df925078f04df.jpg" } ] } ], "index": 13 }, { "bbox": [ 304, 437, 547, 498 ], "type": "text", "angle": 0, "lines": [ { "bbox": [ 304, 437, 547, 498 ], "spans": [ { "bbox": [ 304, 437, 547, 498 ], "type": "text", "content": "In the following, we only consider the case in which objects in " }, { "bbox": [ 304, 437, 547, 498 ], "type": "inline_equation", "content": "S" }, { "bbox": [ 304, 437, 547, 498 ], "type": "text", "content": " are 26-connected and objects in " }, { "bbox": [ 304, 437, 547, 498 ], "type": "inline_equation", "content": "\\overline{S}" }, { "bbox": [ 304, 437, 547, 498 ], "type": "text", "content": " are 6-connected. To calculate the genus of a 26-connected " }, { "bbox": [ 304, 437, 547, 498 ], "type": "inline_equation", "content": "SG_{26}(S)" }, { "bbox": [ 304, 437, 547, 498 ], "type": "text", "content": " we can derive " }, { "bbox": [ 304, 437, 547, 498 ], "type": "inline_equation", "content": "G_{6}(\\overline{S})" }, { "bbox": [ 304, 437, 547, 498 ], "type": "text", "content": " using Equation 2 or 3 and use the following relation [14]:" } ] } ], "index": 14 }, { "bbox": [ 380, 503, 545, 517 ], "type": "interline_equation", "angle": 0, "lines": [ { "bbox": [ 380, 503, 545, 517 ], "spans": [ { "bbox": [ 380, 503, 545, 517 ], "type": "interline_equation", "content": "G _ {2 6} (S) = G _ {6} (\\bar {S}) - 1 \\tag {4}", "image_path": "772df1632562224a103e50ef4c8407f5615caf32bc342293186b2b727062f62d.jpg" } ] } ], "index": 15 }, { "bbox": [ 304, 523, 547, 619 ], "type": "text", "angle": 0, "lines": [ { "bbox": [ 304, 523, 547, 619 ], "spans": [ { "bbox": [ 304, 523, 547, 619 ], "type": "text", "content": "Simple points Crucial to the skeletonization of digital images is the definition of a simple point. A point belonging to " }, { "bbox": [ 304, 523, 547, 619 ], "type": "inline_equation", "content": "S" }, { "bbox": [ 304, 523, 547, 619 ], "type": "text", "content": " is simple if it can be deleted, that is changed from 1 to 0, without altering the image's topology. Morgenthaler [21] shows that this is the case if the deletion of a point does not change the number of objects and holes in " }, { "bbox": [ 304, 523, 547, 619 ], "type": "inline_equation", "content": "S" }, { "bbox": [ 304, 523, 547, 619 ], "type": "text", "content": " and " }, { "bbox": [ 304, 523, 547, 619 ], "type": "inline_equation", "content": "\\overline{S}" }, { "bbox": [ 304, 523, 547, 619 ], "type": "text", "content": ". Using " }, { "bbox": [ 304, 523, 547, 619 ], "type": "inline_equation", "content": "\\delta" }, { "bbox": [ 304, 523, 547, 619 ], "type": "text", "content": " to denote the difference in topology between " }, { "bbox": [ 304, 523, 547, 619 ], "type": "inline_equation", "content": "S" }, { "bbox": [ 304, 523, 547, 619 ], "type": "text", "content": " and " }, { "bbox": [ 304, 523, 547, 619 ], "type": "inline_equation", "content": "S \\setminus \\{p\\}" }, { "bbox": [ 304, 523, 547, 619 ], "type": "text", "content": ", we can write this relation as:" } ] } ], "index": 16 }, { "bbox": [ 312, 624, 545, 654 ], "type": "interline_equation", "angle": 0, "lines": [ { "bbox": [ 312, 624, 545, 654 ], "spans": [ { "bbox": [ 312, 624, 545, 654 ], "type": "interline_equation", "content": "\\begin{array}{r l} p \\text {i s s i m p l e} & \\Longleftrightarrow \\\\ \\delta O (S) = 0, \\delta O (\\bar {S}) = 0, \\delta H (S) = 0, \\delta H (\\bar {S}) = 0. \\end{array} \\tag {5}", "image_path": "abb5596fb3d13c1f11ddbc52ea55e3defcf4f4875bf88d00745acd6f17b9f80f.jpg" } ] } ], "index": 17 }, { "bbox": [ 304, 658, 547, 695 ], "type": "text", "angle": 0, "lines": [ { "bbox": [ 304, 658, 547, 695 ], "spans": [ { "bbox": [ 304, 658, 547, 695 ], "type": "text", "content": "Lee et al. [15] prove that these conditions are equivalent to calculating the change in the number of objects and Euler characteristic of " }, { "bbox": [ 304, 658, 547, 695 ], "type": "inline_equation", "content": "S" }, { "bbox": [ 304, 658, 547, 695 ], "type": "text", "content": " in a point's local 26-neighborhood:" } ] } ], "index": 18 }, { "bbox": [ 332, 700, 545, 715 ], "type": "interline_equation", "angle": 0, "lines": [ { "bbox": [ 332, 700, 545, 715 ], "spans": [ { "bbox": [ 332, 700, 545, 715 ], "type": "interline_equation", "content": "p \\text {i s} \\quad \\Longleftrightarrow \\delta O (S) = 0, \\delta G _ {2 6} (S) = 0. \\tag {6}", "image_path": "f826989a9d6b1c821ec3c60bb57af92df0e79b01561663a576882c7dd5f46c7d.jpg" } ] } ], "index": 19 } ], "discarded_blocks": [ { "bbox": [ 293, 748, 318, 758 ], "type": "page_number", "angle": 0, "lines": [ { "bbox": [ 293, 748, 318, 758 ], "spans": [ { "bbox": [ 293, 748, 318, 758 ], "type": "text", "content": "21395" } ] } ], "index": 20 } ], "page_size": [ 612, 792 ], "page_idx": 1 }, { "para_blocks": [ { "bbox": [ 47, 71, 103, 83 ], "type": "title", "angle": 0, "lines": [ { "bbox": [ 47, 71, 103, 83 ], "spans": [ { "bbox": [ 47, 71, 103, 83 ], "type": "text", "content": "3. Method" } ] } ], "index": 0 }, { "bbox": [ 46, 91, 289, 199 ], "type": "text", "angle": 0, "lines": [ { "bbox": [ 46, 91, 289, 199 ], "spans": [ { "bbox": [ 46, 91, 289, 199 ], "type": "text", "content": "Arguably, the most common class of skeletonization algorithms for digital images are iterative boundary-peeling methods [32]. These algorithms are based on the repeated removal of simple points until only the skeleton remains. [4, 15, 17, 21]. At its core, our skeletonization algorithm follows the same paradigm (see Figure 2). To ensure that our method is compatible with gradient-based optimization while remaining topology-preserving, we make the following contributions:" } ] } ], "index": 1 }, { "bbox": [ 57, 206, 289, 434 ], "type": "list", "angle": 0, "index": 6, "blocks": [ { "bbox": [ 57, 206, 287, 266 ], "type": "text", "angle": 0, "lines": [ { "bbox": [ 57, 206, 287, 266 ], "spans": [ { "bbox": [ 57, 206, 287, 266 ], "type": "text", "content": "- We introduce two methods to differentiably identify simple points (see Section 3.1). One solution relies on the calculation of the Euler characteristic, and the other one is based on a set of Boolean rules that evaluate a point's 26-neighborhood." } ] } ], "index": 2 }, { "bbox": [ 57, 270, 288, 306 ], "type": "text", "angle": 0, "lines": [ { "bbox": [ 57, 270, 288, 306 ], "spans": [ { "bbox": [ 57, 270, 288, 306 ], "type": "text", "content": "- We adopt a scheme to safely delete multiple simple points at once, enabling the parallelization of our algorithm (see Section 3.2)." } ] } ], "index": 3 }, { "bbox": [ 57, 310, 288, 358 ], "type": "text", "angle": 0, "lines": [ { "bbox": [ 57, 310, 288, 358 ], "spans": [ { "bbox": [ 57, 310, 288, 358 ], "type": "text", "content": "- We introduce a strategy to apply our algorithm to nonbinary inputs and integrate it with gradient-based optimization by employing the reparametrization trick and a straight-through estimator (see Section 3.3)." } ] } ], "index": 4 }, { "bbox": [ 57, 361, 289, 434 ], "type": "text", "angle": 0, "lines": [ { "bbox": [ 57, 361, 289, 434 ], "spans": [ { "bbox": [ 57, 361, 289, 434 ], "type": "text", "content": "- All of above's contributions are formulated using matrix additions and multiplications, convolutional operations, basic non-linear functions, and sampling from a uniform probability distribution. We combine them into a single PyTorch module, which we make publicly available (see Section 3.4)." } ] } ], "index": 5 } ], "sub_type": "text" }, { "type": "image", "bbox": [ 56, 450, 279, 541 ], "blocks": [ { "bbox": [ 56, 450, 279, 541 ], "lines": [ { "bbox": [ 56, 450, 279, 541 ], "spans": [ { "bbox": [ 56, 450, 279, 541 ], "type": "image", "image_path": "14cca7ba614427da47a0ef596baf812bf3e71d9ae249e7d9f72700b09b0fa7f5.jpg" } ] } ], "index": 7, "angle": 0, "type": "image_body" }, { "bbox": [ 46, 544, 288, 599 ], "lines": [ { "bbox": [ 46, 544, 288, 599 ], "spans": [ { "bbox": [ 46, 544, 288, 599 ], "type": "text", "content": "Figure 2. Data flow through an iterative boundary-peeling skeletonization algorithm. Our method follows the same paradigm, while ensuring that the identification of simple points, endpoints, and the subfield-based parallelization are all compatible with gradient-based optimization." } ] } ], "index": 8, "angle": 0, "type": "image_caption" } ], "index": 7 }, { "bbox": [ 47, 619, 209, 632 ], "type": "title", "angle": 0, "lines": [ { "bbox": [ 47, 619, 209, 632 ], "spans": [ { "bbox": [ 47, 619, 209, 632 ], "type": "text", "content": "3.1. Identification of simple points" } ] } ], "index": 9 }, { "bbox": [ 47, 637, 270, 650 ], "type": "title", "angle": 0, "lines": [ { "bbox": [ 47, 637, 270, 650 ], "spans": [ { "bbox": [ 47, 637, 270, 650 ], "type": "text", "content": "3.1.1 Euler characteristic to identify simple points" } ] } ], "index": 10 }, { "bbox": [ 46, 656, 287, 692 ], "type": "text", "angle": 0, "lines": [ { "bbox": [ 46, 656, 287, 692 ], "spans": [ { "bbox": [ 46, 656, 287, 692 ], "type": "text", "content": "Lobregt et al. [17] base their detection of simple points on the observation that the removal of a simple point does not alter the genus of its 26-neighborhood:" } ] } ], "index": 11 }, { "bbox": [ 100, 701, 287, 714 ], "type": "interline_equation", "angle": 0, "lines": [ { "bbox": [ 100, 701, 287, 714 ], "spans": [ { "bbox": [ 100, 701, 287, 714 ], "type": "interline_equation", "content": "p \\text {i s} \\quad \\text {s i m p l e} \\Rightarrow \\delta G _ {2 6} (S) = 0, \\tag {7}", "image_path": "d1f28fe5d9bf7222a320078f7de96d3f0db35eb10aaeb8334940c45a665f6c05.jpg" } ] } ], "index": 12 }, { "bbox": [ 304, 72, 545, 143 ], "type": "text", "angle": 0, "lines": [ { "bbox": [ 304, 72, 545, 143 ], "spans": [ { "bbox": [ 304, 72, 545, 143 ], "type": "text", "content": "which is a relaxation of Equation 6. To efficiently determine " }, { "bbox": [ 304, 72, 545, 143 ], "type": "inline_equation", "content": "\\delta G_{26}(\\mathbf{S})" }, { "bbox": [ 304, 72, 545, 143 ], "type": "text", "content": ", their algorithm assesses the change of the genus in each of the 26-neighborhood's eight octants and sums their contributions. Thereby, they rely on a look-up table in which each entry corresponds to one of the " }, { "bbox": [ 304, 72, 545, 143 ], "type": "inline_equation", "content": "2^{8}" }, { "bbox": [ 304, 72, 545, 143 ], "type": "text", "content": " possible configurations of an octant." } ] } ], "index": 13 }, { "bbox": [ 304, 144, 547, 276 ], "type": "text", "angle": 0, "lines": [ { "bbox": [ 304, 144, 547, 276 ], "spans": [ { "bbox": [ 304, 144, 547, 276 ], "type": "text", "content": "In order to reduce the number of comparisons and provide a smoother learning signal for backpropagation, we use eight convolutions with pre-defined kernels to determine the number of vertices, edges, faces, and octants (see Figure 3). By inserting these into Equation 3, we calculate " }, { "bbox": [ 304, 144, 547, 276 ], "type": "inline_equation", "content": "G_{6}(\\overline{S})" }, { "bbox": [ 304, 144, 547, 276 ], "type": "text", "content": " of each 26-neighborhood. Afterwards, we repeat this process with the central point of each neighborhood set to zero and assess whether the Euler characteristic has changed. This process is parallelized while ensuring that only one point in each 26-neighborhood is deleted at a given time. To this end, we use multiple sets of points given by" } ] } ], "index": 14 }, { "bbox": [ 330, 285, 545, 304 ], "type": "interline_equation", "angle": 0, "lines": [ { "bbox": [ 330, 285, 545, 304 ], "spans": [ { "bbox": [ 330, 285, 545, 304 ], "type": "interline_equation", "content": "S _ {i, j, k} \\in \\left\\{p ^ {\\prime} (x + i, y + j, z + k) \\right\\}; \\tag {8}", "image_path": "82ef6133017e4269a2c350791ce88882e67bd2d590da32131b608d2094bfc1e7.jpg" } ] } ], "index": 15 }, { "bbox": [ 364, 300, 518, 312 ], "type": "interline_equation", "angle": 0, "lines": [ { "bbox": [ 364, 300, 518, 312 ], "spans": [ { "bbox": [ 364, 300, 518, 312 ], "type": "interline_equation", "content": "x, y, z \\in \\{0, 2, 4, \\dots \\}, i, j, k \\in \\{0, 1 \\}", "image_path": "651d86cb4f09e3a57726719d966f10ef3421685db92b74f00e922d557e5235ac.jpg" } ] } ], "index": 16 }, { "bbox": [ 306, 322, 547, 369 ], "type": "text", "angle": 0, "lines": [ { "bbox": [ 306, 322, 547, 369 ], "spans": [ { "bbox": [ 306, 322, 547, 369 ], "type": "text", "content": "Cycling through all combinations of " }, { "bbox": [ 306, 322, 547, 369 ], "type": "inline_equation", "content": "i, j" }, { "bbox": [ 306, 322, 547, 369 ], "type": "text", "content": ", and " }, { "bbox": [ 306, 322, 547, 369 ], "type": "inline_equation", "content": "k" }, { "bbox": [ 306, 322, 547, 369 ], "type": "text", "content": " yields eight subfields of points that can be processed simultaneously. The same subfields are also used during the later removal of simple points (see Section 3.2)." } ] } ], "index": 17 }, { "bbox": [ 304, 370, 547, 466 ], "type": "text", "angle": 0, "lines": [ { "bbox": [ 304, 370, 547, 466 ], "spans": [ { "bbox": [ 304, 370, 547, 466 ], "type": "text", "content": "Lee et al. [15] show that the invariance of the Euler characteristic under deletion of a point is a necessary but not sufficient condition for it being simple (cf. Equation 6). Consequently, above's strategy slightly overestimates the number of simple points [15]. On the set of all possible " }, { "bbox": [ 304, 370, 547, 466 ], "type": "inline_equation", "content": "2^{26}" }, { "bbox": [ 304, 370, 547, 466 ], "type": "text", "content": " configurations of a 26-neighborhood, above's algorithm characterizes the central point as simple in " }, { "bbox": [ 304, 370, 547, 466 ], "type": "inline_equation", "content": "40.07\\%" }, { "bbox": [ 304, 370, 547, 466 ], "type": "text", "content": " of cases when in fact only " }, { "bbox": [ 304, 370, 547, 466 ], "type": "inline_equation", "content": "38.72\\%" }, { "bbox": [ 304, 370, 547, 466 ], "type": "text", "content": " are truly simple." } ] } ], "index": 18 }, { "bbox": [ 306, 481, 515, 493 ], "type": "title", "angle": 0, "lines": [ { "bbox": [ 306, 481, 515, 493 ], "spans": [ { "bbox": [ 306, 481, 515, 493 ], "type": "text", "content": "3.1.2 Boolean characterization of simple points" } ] } ], "index": 19 }, { "bbox": [ 304, 500, 547, 548 ], "type": "text", "angle": 0, "lines": [ { "bbox": [ 304, 500, 547, 548 ], "spans": [ { "bbox": [ 304, 500, 547, 548 ], "type": "text", "content": "For this reason, we propose a second method that identifies the exact set of simple points. It is based on work by Bertrand et al. [4] who introduce the following Boolean characterization of a simple point:" } ] } ], "index": 20 }, { "bbox": [ 332, 557, 515, 571 ], "type": "interline_equation", "angle": 0, "lines": [ { "bbox": [ 332, 557, 515, 571 ], "spans": [ { "bbox": [ 332, 557, 515, 571 ], "type": "interline_equation", "content": "p \\text {i s} \\quad \\Longleftrightarrow \\quad (\\# \\overline {{X _ {6}}} = 1) o r (\\# X _ {2 6} = 1)", "image_path": "4d228baadde81646617406cafb58cbb7e2b1dcc5f2bede7eca661aa26d65dbd0.jpg" } ] } ], "index": 21 }, { "bbox": [ 339, 573, 444, 585 ], "type": "interline_equation", "angle": 0, "lines": [ { "bbox": [ 339, 573, 444, 585 ], "spans": [ { "bbox": [ 339, 573, 444, 585 ], "type": "interline_equation", "content": "o r \\left(\\# B _ {2 6} = 0, \\# X _ {1 8} = 1\\right)", "image_path": "f4a394f1505a8c5948bdc31540d1747332b6871be2f48a46aa2e217a3686a510.jpg" } ] } ], "index": 22 }, { "bbox": [ 339, 583, 545, 601 ], "type": "interline_equation", "angle": 0, "lines": [ { "bbox": [ 339, 583, 545, 601 ], "spans": [ { "bbox": [ 339, 583, 545, 601 ], "type": "interline_equation", "content": "o r \\left(\\# \\overline {{A _ {6}}} = 0, \\# B _ {2 6} = 0, \\# B _ {1 8} = 0, \\right. \\tag {9}", "image_path": "e4f689a4e012a6cd00d8e6b46d1d4ce2c7bc9d2380b63e3678837a45b7abfaba.jpg" } ] } ], "index": 23 }, { "bbox": [ 356, 605, 463, 617 ], "type": "interline_equation", "angle": 0, "lines": [ { "bbox": [ 356, 605, 463, 617 ], "spans": [ { "bbox": [ 356, 605, 463, 617 ], "type": "interline_equation", "content": "\\# \\overline {{X _ {6}}} - \\# \\overline {{A _ {1 8}}} + \\# \\overline {{A _ {2 6}}} = 1)", "image_path": "e4fc7852ecb0ecdb28093378041b2f73febaa57cee044fbd7711657166b815a8.jpg" } ] } ], "index": 24 }, { "bbox": [ 306, 629, 545, 677 ], "type": "text", "angle": 0, "lines": [ { "bbox": [ 306, 629, 545, 677 ], "spans": [ { "bbox": [ 306, 629, 545, 677 ], "type": "text", "content": "where " }, { "bbox": [ 306, 629, 545, 677 ], "type": "inline_equation", "content": "\\# X_{n}" }, { "bbox": [ 306, 629, 545, 677 ], "type": "text", "content": " and " }, { "bbox": [ 306, 629, 545, 677 ], "type": "inline_equation", "content": "\\# \\overline{X_n}" }, { "bbox": [ 306, 629, 545, 677 ], "type": "text", "content": " are the number of " }, { "bbox": [ 306, 629, 545, 677 ], "type": "inline_equation", "content": "n" }, { "bbox": [ 306, 629, 545, 677 ], "type": "text", "content": "-neighbors of a point belonging to " }, { "bbox": [ 306, 629, 545, 677 ], "type": "inline_equation", "content": "S" }, { "bbox": [ 306, 629, 545, 677 ], "type": "text", "content": " and " }, { "bbox": [ 306, 629, 545, 677 ], "type": "inline_equation", "content": "\\overline{S}" }, { "bbox": [ 306, 629, 545, 677 ], "type": "text", "content": ", respectively. " }, { "bbox": [ 306, 629, 545, 677 ], "type": "inline_equation", "content": "\\# B_{26}, \\# \\overline{A_6}, \\# B_{18}, \\# \\overline{A_{18}}" }, { "bbox": [ 306, 629, 545, 677 ], "type": "text", "content": ", and " }, { "bbox": [ 306, 629, 545, 677 ], "type": "inline_equation", "content": "\\# \\overline{A_{26}}" }, { "bbox": [ 306, 629, 545, 677 ], "type": "text", "content": " correspond to the number of specific cell configurations depicted in Figure 4." } ] } ], "index": 25 }, { "bbox": [ 304, 677, 547, 713 ], "type": "text", "angle": 0, "lines": [ { "bbox": [ 304, 677, 547, 713 ], "spans": [ { "bbox": [ 304, 677, 547, 713 ], "type": "text", "content": "Similar as before, the presence of these five configurations and their 6-, 8-, and 12-rotational equivalents can be efficiently checked by convolving the image with pre-defined" } ] } ], "index": 26 } ], "discarded_blocks": [ { "bbox": [ 293, 748, 318, 757 ], "type": "page_number", "angle": 0, "lines": [ { "bbox": [ 293, 748, 318, 757 ], "spans": [ { "bbox": [ 293, 748, 318, 757 ], "type": "text", "content": "21396" } ] } ], "index": 27 } ], "page_size": [ 612, 792 ], "page_idx": 2 }, { "para_blocks": [ { "type": "image", "bbox": [ 61, 70, 532, 221 ], "blocks": [ { "bbox": [ 61, 70, 532, 221 ], "lines": [ { "bbox": [ 61, 70, 532, 221 ], "spans": [ { "bbox": [ 61, 70, 532, 221 ], "type": "image", "image_path": "efa51282cbb9e0c3237bb33992e68ef9cf4ec0a01a688b5b72996675f7c1c8cc.jpg" } ] } ], "index": 0, "angle": 0, "type": "image_body" }, { "bbox": [ 46, 225, 547, 258 ], "lines": [ { "bbox": [ 46, 225, 547, 258 ], "spans": [ { "bbox": [ 46, 225, 547, 258 ], "type": "text", "content": "Figure 3. In order to calculate the Euler characteristic of a point, we initially invert its 26-neighborhood. Next, we determine the number of vertices, edges, faces, and octants of the background via eight simple convolutions with predefined kernels. Finally, Equations 3 and 4 are used to derive the Euler characteristic of the foreground." } ] } ], "index": 1, "angle": 0, "type": "image_caption" } ], "index": 0 }, { "bbox": [ 46, 278, 288, 339 ], "type": "text", "angle": 0, "lines": [ { "bbox": [ 46, 278, 288, 339 ], "spans": [ { "bbox": [ 46, 278, 288, 339 ], "type": "text", "content": "kernels. Compared to our first strategy, this algorithm trades off computational efficiency for accuracy. It requires a total of 57 convolutions with three-dimensional kernels, but is guaranteed to correctly classify all points of a binary image as either simple or not." } ] } ], "index": 2 }, { "type": "image", "bbox": [ 65, 351, 271, 427 ], "blocks": [ { "bbox": [ 65, 351, 271, 427 ], "lines": [ { "bbox": [ 65, 351, 271, 427 ], "spans": [ { "bbox": [ 65, 351, 271, 427 ], "type": "image", "image_path": "e61f785649ef9ad63c50f88028065d51eb964a27244e9780d157f7ab6e23eef4.jpg" } ] } ], "index": 3, "angle": 0, "type": "image_body" } ], "index": 3 }, { "type": "image", "bbox": [ 100, 427, 236, 503 ], "blocks": [ { "bbox": [ 100, 427, 236, 503 ], "lines": [ { "bbox": [ 100, 427, 236, 503 ], "spans": [ { "bbox": [ 100, 427, 236, 503 ], "type": "image", "image_path": "942161df44562202f8a28e168ec1bdf95bc24c1c9cb8f2de87d87da0bc90f0c5.jpg" } ] } ], "index": 4, "angle": 0, "type": "image_body" }, { "bbox": [ 47, 509, 288, 532 ], "lines": [ { "bbox": [ 47, 509, 288, 532 ], "spans": [ { "bbox": [ 47, 509, 288, 532 ], "type": "text", "content": "Figure 4. The five cell configuration introduced by Bertrand et al. [4] used for Boolean characterization of simple points." } ] } ], "index": 5, "angle": 0, "type": "image_caption" } ], "index": 4 }, { "bbox": [ 47, 551, 253, 564 ], "type": "title", "angle": 0, "lines": [ { "bbox": [ 47, 551, 253, 564 ], "spans": [ { "bbox": [ 47, 551, 253, 564 ], "type": "text", "content": "3.2. Parallelization and endpoint conditions" } ] } ], "index": 6 }, { "bbox": [ 46, 570, 289, 714 ], "type": "text", "angle": 0, "lines": [ { "bbox": [ 46, 570, 289, 714 ], "spans": [ { "bbox": [ 46, 570, 289, 714 ], "type": "text", "content": "Sequentially checking each point using above's conditions and deleting them if they are simple already constitutes a functioning skeletonization algorithm. However, this strategy is very inefficient when applied to large three-dimensional images. Naively deleting all simple points at once is not possible as simultaneous removal of neighboring points may affect the object's topology even if both points are simple. For this reason, previous works have researched strategies to safely remove multiple simple points in parallel [5, 15, 18, 21, 31, 42]. We adopt a subiterative scheme based on the same eight subfields that are already used during the calculation of the Euler characteristic (see Equation 8) [5]." } ] } ], "index": 7 }, { "bbox": [ 304, 278, 547, 361 ], "type": "text", "angle": 0, "lines": [ { "bbox": [ 304, 278, 547, 361 ], "spans": [ { "bbox": [ 304, 278, 547, 361 ], "type": "text", "content": "In conventional skeletonization algorithms, the program terminates once a full iteration does not result in any points being deleted. To keep the number of operations comparable during repeated application, we explicitly provide the number of outer-loop iterations. This simple scalar hyperparameter can be easily tuned on a few representative samples of the considered dataset." } ] } ], "index": 8 }, { "bbox": [ 304, 363, 547, 460 ], "type": "text", "angle": 0, "lines": [ { "bbox": [ 304, 363, 547, 460 ], "spans": [ { "bbox": [ 304, 363, 547, 460 ], "type": "text", "content": "Merely preserving non-simple points would lead to a topological skeleton. For example, a solid object without any holes or cavities would be reduced to a single point. For many applications in image processing, it is desirable to also preserve some information about the image's geometry, such as the existence and position of extremities. This can be achieved by also preserving so-called endpoints. Our algorithm uses the following definition of an endpoint:" } ] } ], "index": 9 }, { "bbox": [ 361, 470, 545, 483 ], "type": "interline_equation", "angle": 0, "lines": [ { "bbox": [ 361, 470, 545, 483 ], "spans": [ { "bbox": [ 361, 470, 545, 483 ], "type": "interline_equation", "content": "p \\text {i s} \\quad \\Longleftrightarrow \\quad \\# X _ {2 6} \\leq 1 \\tag {10}", "image_path": "20c094b12e9a50c5b4acd3a21104330aab473c7e4e713283e87ed8fe5548b8b8.jpg" } ] } ], "index": 10 }, { "bbox": [ 304, 493, 547, 565 ], "type": "text", "angle": 0, "lines": [ { "bbox": [ 304, 493, 547, 565 ], "spans": [ { "bbox": [ 304, 493, 547, 565 ], "type": "text", "content": "Other definitions for endpoints could potentially be integrated with our algorithm and would result in different properties of the obtained skeleton. For example, endpoint conditions could be chosen to extract a medial surface instead of a medial axis, or the number of short extremities, sometimes called spurs, may be reduced [34, 46]." } ] } ], "index": 11 }, { "bbox": [ 306, 574, 476, 588 ], "type": "title", "angle": 0, "lines": [ { "bbox": [ 306, 574, 476, 588 ], "spans": [ { "bbox": [ 306, 574, 476, 588 ], "type": "text", "content": "3.3. Processing of continuous inputs" } ] } ], "index": 12 }, { "bbox": [ 304, 594, 547, 714 ], "type": "text", "angle": 0, "lines": [ { "bbox": [ 304, 594, 547, 714 ], "spans": [ { "bbox": [ 304, 594, 547, 714 ], "type": "text", "content": "The previously introduced definitions for simple points and endpoints are only valid for binary images. However, in many applications the input is often a matrix of continuous values, such as probability maps output by a learning algorithm. Simply rounding these inputs inhibits learning via backpropagation. We circumvent this issue by treating each point as a discrete, stochastic node modeled by a Bernoulli distribution, and use the reparametrization trick and a straight-through estimator to facilitate sampling and gradient estimation from it [3, 12, 19]." } ] } ], "index": 13 } ], "discarded_blocks": [ { "bbox": [ 293, 748, 317, 757 ], "type": "page_number", "angle": 0, "lines": [ { "bbox": [ 293, 748, 317, 757 ], "spans": [ { "bbox": [ 293, 748, 317, 757 ], "type": "text", "content": "21397" } ] } ], "index": 14 } ], "page_size": [ 612, 792 ], "page_idx": 3 }, { "para_blocks": [ { "type": "image", "bbox": [ 49, 70, 545, 268 ], "blocks": [ { "bbox": [ 49, 70, 545, 268 ], "lines": [ { "bbox": [ 49, 70, 545, 268 ], "spans": [ { "bbox": [ 49, 70, 545, 268 ], "type": "image", "image_path": "2553cf69eab596650fa0c2b7fe38fe84e828fd78670978cd383c7527fd7e197a.jpg" } ] } ], "index": 0, "angle": 0, "type": "image_body" }, { "bbox": [ 46, 271, 547, 307 ], "lines": [ { "bbox": [ 46, 271, 547, 307 ], "spans": [ { "bbox": [ 46, 271, 547, 307 ], "type": "text", "content": "Figure 5. The results of applying the seven tested skeletonization algorithm to representative samples of three diverse datasets. Of the six algorithms that are compatible with gradient-based optimization, only our two methods are able to extract a thin, topology-preserving skeleton, similar to the one obtained using the non-differential baseline. Additional samples are shown in the Supplementary Material." } ] } ], "index": 1, "angle": 0, "type": "image_caption" } ], "index": 0 }, { "bbox": [ 46, 325, 288, 361 ], "type": "text", "angle": 0, "lines": [ { "bbox": [ 46, 325, 288, 361 ], "spans": [ { "bbox": [ 46, 325, 288, 361 ], "type": "text", "content": "The reparametrization trick splits each node into a differentiable function, the raw grayscale input, and a fixed noise distribution [11, 12, 19]. We can sample from each node via" } ] } ], "index": 2 }, { "bbox": [ 70, 369, 287, 397 ], "type": "interline_equation", "angle": 0, "lines": [ { "bbox": [ 70, 369, 287, 397 ], "spans": [ { "bbox": [ 70, 369, 287, 397 ], "type": "interline_equation", "content": "X = \\sigma \\left(\\frac {(\\log \\alpha + \\beta L)}{\\tau}\\right), \\alpha = \\frac {\\alpha_ {1}}{(1 - \\alpha_ {1})}, \\tag {11}", "image_path": "5e5db8284cf27a0a6ada5aebf43eb08526843ea6aa2b2361aab3f124d889bea7.jpg" } ] } ], "index": 3 }, { "bbox": [ 46, 405, 287, 501 ], "type": "text", "angle": 0, "lines": [ { "bbox": [ 46, 405, 287, 501 ], "spans": [ { "bbox": [ 46, 405, 287, 501 ], "type": "text", "content": "where " }, { "bbox": [ 46, 405, 287, 501 ], "type": "inline_equation", "content": "\\alpha_{1} \\in (0, 1)" }, { "bbox": [ 46, 405, 287, 501 ], "type": "text", "content": " is the probability of the input being " }, { "bbox": [ 46, 405, 287, 501 ], "type": "inline_equation", "content": "1" }, { "bbox": [ 46, 405, 287, 501 ], "type": "text", "content": ", " }, { "bbox": [ 46, 405, 287, 501 ], "type": "inline_equation", "content": "\\sigma" }, { "bbox": [ 46, 405, 287, 501 ], "type": "text", "content": " denotes the sigmoid function, " }, { "bbox": [ 46, 405, 287, 501 ], "type": "inline_equation", "content": "L" }, { "bbox": [ 46, 405, 287, 501 ], "type": "text", "content": " is a sample from a Logistic distribution that is scaled by factor " }, { "bbox": [ 46, 405, 287, 501 ], "type": "inline_equation", "content": "\\beta \\in [0, \\infty)" }, { "bbox": [ 46, 405, 287, 501 ], "type": "text", "content": ", and " }, { "bbox": [ 46, 405, 287, 501 ], "type": "inline_equation", "content": "\\tau \\in (0, \\infty)" }, { "bbox": [ 46, 405, 287, 501 ], "type": "text", "content": " is the Boltzmann temperature. Both " }, { "bbox": [ 46, 405, 287, 501 ], "type": "inline_equation", "content": "\\beta" }, { "bbox": [ 46, 405, 287, 501 ], "type": "text", "content": " and " }, { "bbox": [ 46, 405, 287, 501 ], "type": "inline_equation", "content": "\\tau" }, { "bbox": [ 46, 405, 287, 501 ], "type": "text", "content": " control the entropy of the distribution. Others have proposed gradually annealing these parameters as learning progresses or even updating them via backpropagation [11, 12, 19]. In this work, we treat them as simple tunable hyperparameter." } ] } ], "index": 4 }, { "bbox": [ 46, 502, 288, 573 ], "type": "text", "angle": 0, "lines": [ { "bbox": [ 46, 502, 288, 573 ], "spans": [ { "bbox": [ 46, 502, 288, 573 ], "type": "text", "content": "Afterwards, we discretize the obtained samples using a straight-through gradient estimator [3, 29]. It returns the rounded binary value during the forward pass. Instead of using the zero gradient of the rounding operation during the backward pass, the modified chain rule is applied and the identity function is used as proxy gradient." } ] } ], "index": 5 }, { "bbox": [ 47, 582, 200, 594 ], "type": "title", "angle": 0, "lines": [ { "bbox": [ 47, 582, 200, 594 ], "spans": [ { "bbox": [ 47, 582, 200, 594 ], "type": "text", "content": "3.4. Implementation in PyTorch" } ] } ], "index": 6 }, { "bbox": [ 46, 601, 288, 685 ], "type": "text", "angle": 0, "lines": [ { "bbox": [ 46, 601, 288, 685 ], "spans": [ { "bbox": [ 46, 601, 288, 685 ], "type": "text", "content": "Our proposed algorithm consists exclusively of matrix additions, multiplications, convolutional operations, basic activation functions, and sampling from a uniform probability distribution. As such, it can easily be implemented in any major deep learning library and runs efficiently on graphics processing units. Our skeletonization module, which we make publicly available1, is implemented in PyTorch" } ] } ], "index": 7 }, { "bbox": [ 306, 324, 545, 349 ], "type": "text", "angle": 0, "lines": [ { "bbox": [ 306, 324, 545, 349 ], "spans": [ { "bbox": [ 306, 324, 545, 349 ], "type": "text", "content": "[27] and is fully integrated with its automatic differentiation engine." } ] } ], "index": 8 }, { "bbox": [ 306, 360, 446, 373 ], "type": "title", "angle": 0, "lines": [ { "bbox": [ 306, 360, 446, 373 ], "spans": [ { "bbox": [ 306, 360, 446, 373 ], "type": "text", "content": "4. Experiments and results" } ] } ], "index": 9 }, { "bbox": [ 304, 380, 547, 475 ], "type": "text", "angle": 0, "lines": [ { "bbox": [ 304, 380, 547, 475 ], "spans": [ { "bbox": [ 304, 380, 547, 475 ], "type": "text", "content": "Initially, we benchmark the performance of our skeletonization algorithm with regard to spatial and topological correctness, run time, and the ability to combine it with backpropagation (see Section 4.1). Afterwards, we showcase the utility of our method by integrating it with two medical image processing pipelines: deep-learning-based blood vessel segmentation and multimodal registration of the mandible (see Section 4.2)." } ] } ], "index": 10 }, { "bbox": [ 317, 475, 528, 487 ], "type": "text", "angle": 0, "lines": [ { "bbox": [ 317, 475, 528, 487 ], "spans": [ { "bbox": [ 317, 475, 528, 487 ], "type": "text", "content": "For the experiments, we use three different datasets:" } ] } ], "index": 11 }, { "bbox": [ 316, 495, 547, 622 ], "type": "list", "angle": 0, "index": 15, "blocks": [ { "bbox": [ 316, 495, 547, 531 ], "type": "text", "angle": 0, "lines": [ { "bbox": [ 316, 495, 547, 531 ], "spans": [ { "bbox": [ 316, 495, 547, 531 ], "type": "text", "content": "- the DRIVE dataset consisting of 40 two-dimensional retinal color fundus photographs and matching annotations of the visible blood vessels [39]," } ] } ], "index": 12 }, { "bbox": [ 316, 535, 547, 583 ], "type": "text", "angle": 0, "lines": [ { "bbox": [ 316, 535, 547, 583 ], "spans": [ { "bbox": [ 316, 535, 547, 583 ], "type": "text", "content": "- the VesSAP dataset comprising 24 three-dimensional light-sheet microscopy images of murine brains after tissue clearing, staining, and labeling of the vascular network, which we split into 2,400 patches, [41]." } ] } ], "index": 13 }, { "bbox": [ 316, 586, 547, 622 ], "type": "text", "angle": 0, "lines": [ { "bbox": [ 316, 586, 547, 622 ], "spans": [ { "bbox": [ 316, 586, 547, 622 ], "type": "text", "content": "- an in-house dataset of 34 matched three-dimensional CT and MR images and manually extracted segmentation masks of the mandible." } ] } ], "index": 14 } ], "sub_type": "text" }, { "bbox": [ 306, 631, 547, 654 ], "type": "text", "angle": 0, "lines": [ { "bbox": [ 306, 631, 547, 654 ], "spans": [ { "bbox": [ 306, 631, 547, 654 ], "type": "text", "content": "Additional information about each dataset and our experimental setup can be found in the Supplementary Material." } ] } ], "index": 16 } ], "discarded_blocks": [ { "bbox": [ 47, 693, 277, 713 ], "type": "page_footnote", "angle": 0, "lines": [ { "bbox": [ 47, 693, 277, 713 ], "spans": [ { "bbox": [ 47, 693, 277, 713 ], "type": "text", "content": "1https://github.com/martinmenten/skeletonization-for-gradient-based-optimization" } ] } ], "index": 17 }, { "bbox": [ 293, 748, 317, 757 ], "type": "page_number", "angle": 0, "lines": [ { "bbox": [ 293, 748, 317, 757 ], "spans": [ { "bbox": [ 293, 748, 317, 757 ], "type": "text", "content": "21398" } ] } ], "index": 18 } ], "page_size": [ 612, 792 ], "page_idx": 4 }, { "para_blocks": [ { "type": "table", "bbox": [ 64, 80, 530, 325 ], "blocks": [ { "bbox": [ 65, 69, 527, 79 ], "lines": [ { "bbox": [ 65, 69, 527, 79 ], "spans": [ { "bbox": [ 65, 69, 527, 79 ], "type": "text", "content": "Table 1. Quantitative comparison of the topological accuracy and run time of seven skeletonization algorithms on three datasets." } ] } ], "index": 0, "angle": 0, "type": "table_caption" }, { "bbox": [ 64, 80, 530, 325 ], "lines": [ { "bbox": [ 64, 80, 530, 325 ], "spans": [ { "bbox": [ 64, 80, 530, 325 ], "type": "table", "html": "
DatasetSkeletonization algorithm# pointsβ0errorβ1errorβ2errorRun time [ms]
DRIVENon-differentiable - Bertrand et al. [6]8316±6180±00±0--
Morphological - Shit et al. [38]9926±6671156±19750±23-19±1
Morphological - Viti et al. [44]11834±976266±6245±19-23±3
Neural network - Panichev et al. [26]10420±9156±313±11-14±1
Neural network - Nguyen [24]10619±80610±518±8-117±1
Ours - Euler characteristic8393±6110±00±0-101±3
Ours - Boolean8393±6110±00±0-540±2
VesSAPNon-differentiable - Bertrand et al. [6]471±2120±00±00±0-
Morphological - Shit et al. [38]1914±809173±793±50±120±1
Morphological - Viti et al. [44]3783±17972±223±180±121±2
Neural network - Panichev et al. [26]423±18264±383±50±116±1
Neural network - Nguyen [24]422±18933±194±60±1115±1
Ours (Euler characteristic)540±2450±10±10±1100±2
Ours (Boolean)540±2430±00±00±0520±26
MandibleNon-differentiable - Bertrand et al. [6]236±370±00±00±0-
Morphological - Shit et al. [38]2698±387131±2413±100±037±1
Morphological - Viti et al. [44]4866±7581±181±260±042±1
Neural network - Panichev et al. [26]Insufficient training data
Neural network - Nguyen [24]Insufficient training data
Ours (Euler characteristic)409±671±11±10±0160±2
Ours (Boolean)405±680±00±00±01081±10
", "image_path": "9b5f6513abf2aae86c4aeadea270c73ec8ed391a88ed18104e7d0587d9b6db08.jpg" } ] } ], "index": 1, "angle": 0, "type": "table_body" } ], "index": 1 }, { "bbox": [ 47, 343, 200, 357 ], "type": "title", "angle": 0, "lines": [ { "bbox": [ 47, 343, 200, 357 ], "spans": [ { "bbox": [ 47, 343, 200, 357 ], "type": "text", "content": "4.1. Benchmarking experiments" } ] } ], "index": 2 }, { "bbox": [ 47, 362, 218, 374 ], "type": "title", "angle": 0, "lines": [ { "bbox": [ 47, 362, 218, 374 ], "spans": [ { "bbox": [ 47, 362, 218, 374 ], "type": "text", "content": "4.1.1 Spatial and topological accuracy" } ] } ], "index": 3 }, { "bbox": [ 46, 380, 287, 404 ], "type": "text", "angle": 0, "lines": [ { "bbox": [ 46, 380, 287, 404 ], "spans": [ { "bbox": [ 46, 380, 287, 404 ], "type": "text", "content": "We compare our two skeletonization algorithms with five baselines:" } ] } ], "index": 4 }, { "bbox": [ 57, 412, 288, 540 ], "type": "list", "angle": 0, "index": 8, "blocks": [ { "bbox": [ 57, 412, 288, 449 ], "type": "text", "angle": 0, "lines": [ { "bbox": [ 57, 412, 288, 449 ], "spans": [ { "bbox": [ 57, 412, 288, 449 ], "type": "text", "content": "- a well-established non-differentiable skeletonization algorithm by Bertrand et al. [6], which has been implemented in the open-source DGtal library [1]," } ] } ], "index": 5 }, { "bbox": [ 57, 452, 288, 487 ], "type": "text", "angle": 0, "lines": [ { "bbox": [ 57, 452, 288, 487 ], "spans": [ { "bbox": [ 57, 452, 288, 487 ], "type": "text", "content": "- two morphological skeletonization algorithm based on repeated opening and erosion proposed by Shit et al. [38] and Viti et al. [44], respectively," } ] } ], "index": 6 }, { "bbox": [ 57, 491, 288, 540 ], "type": "text", "angle": 0, "lines": [ { "bbox": [ 57, 491, 288, 540 ], "spans": [ { "bbox": [ 57, 491, 288, 540 ], "type": "text", "content": "- two neural-network-based methods by Panichev et al. [26] and Nguyen [24], respectively, that each train a encoder-decoder network to output the skeleton of a binary input image." } ] } ], "index": 7 } ], "sub_type": "text" }, { "bbox": [ 46, 546, 287, 688 ], "type": "text", "angle": 0, "lines": [ { "bbox": [ 46, 546, 287, 688 ], "spans": [ { "bbox": [ 46, 546, 287, 688 ], "type": "text", "content": "On all three datasets, both of our proposed algorithms produce continuous, thin skeletons that agree well with the nondifferentiable baseline (see Figure 5). When applying the morphological skeletonization algorithms, we observe that continuous blood vessels are split into many small objects along the medial axis. Similarly, the mandible is broken into small components that are positioned at the medial surface of the input structure. The neural-network-based algorithms also cannot preserve the topology of the vascular network when applied to data from the DRIVE and VesSAP datasets, and completely fail to converge during training on the small mandible dataset." } ] } ], "index": 9 }, { "bbox": [ 47, 689, 288, 713 ], "type": "text", "angle": 0, "lines": [ { "bbox": [ 47, 689, 288, 713 ], "spans": [ { "bbox": [ 47, 689, 288, 713 ], "type": "text", "content": "These observations are corroborated by quantitative measurements (see Table 1). We assess the topological correct-" } ] } ], "index": 10 }, { "bbox": [ 304, 344, 547, 498 ], "type": "text", "angle": 0, "lines": [ { "bbox": [ 304, 344, 547, 498 ], "spans": [ { "bbox": [ 304, 344, 547, 498 ], "type": "text", "content": "ness of all skeletons by evaluating the error of the first three Betti numbers, " }, { "bbox": [ 304, 344, 547, 498 ], "type": "inline_equation", "content": "\\beta_0, \\beta_1" }, { "bbox": [ 304, 344, 547, 498 ], "type": "text", "content": ", and " }, { "bbox": [ 304, 344, 547, 498 ], "type": "inline_equation", "content": "\\beta_2" }, { "bbox": [ 304, 344, 547, 498 ], "type": "text", "content": ". These measure the absolute difference of the number of objects, holes, and cavities, respectively, between the input structure and obtained skeleton. Our skeletonization algorithm based on the Boolean characterization of simple points preserves the exact topology of the base object in all cases as does the non-differentiable baseline. The skeletons obtained by the morphological and neural-network-based algorithms both contain topological errors in all three Betti numbers. Furthermore, their produced skeletons are often thicker than one voxel. This is reflected by the substantially larger number of points included in the results." } ] } ], "index": 11 }, { "bbox": [ 306, 515, 416, 528 ], "type": "title", "angle": 0, "lines": [ { "bbox": [ 306, 515, 416, 528 ], "spans": [ { "bbox": [ 306, 515, 416, 528 ], "type": "text", "content": "4.1.2 Run time analysis" } ] } ], "index": 12 }, { "bbox": [ 304, 534, 547, 713 ], "type": "text", "angle": 0, "lines": [ { "bbox": [ 304, 534, 547, 713 ], "spans": [ { "bbox": [ 304, 534, 547, 713 ], "type": "text", "content": "Table 1 also lists the average run time of each skeletonization algorithm when processing images of varying sizes and dimensions. We report the duration of a single forward and backward pass through each skeletonization module as required during gradient-based optimization. All measurements were conducted using a workstation equipped with a Nvidia RTX A6000 GPU (Nvidia Corporation, Santa Clara, California, United States), 128 GB of random access memory, and a AMD Ryzen Threadripper 3970X 32-core central processing unit (Advanced Micro Devices, Inc., Santa Clara, California, United States). We find that our algorithms are slower than both morphological and neural-network-based methods. Still, all algorithms run in a second or less and are fast enough to be effectively employed for the applications described in the following (see Section 4.2)." } ] } ], "index": 13 } ], "discarded_blocks": [ { "bbox": [ 293, 748, 318, 757 ], "type": "page_number", "angle": 0, "lines": [ { "bbox": [ 293, 748, 318, 757 ], "spans": [ { "bbox": [ 293, 748, 318, 757 ], "type": "text", "content": "21399" } ] } ], "index": 14 } ], "page_size": [ 612, 792 ], "page_idx": 5 }, { "para_blocks": [ { "bbox": [ 47, 72, 284, 85 ], "type": "title", "angle": 0, "lines": [ { "bbox": [ 47, 72, 284, 85 ], "spans": [ { "bbox": [ 47, 72, 284, 85 ], "type": "text", "content": "4.1.3 Compatibility with gradient-based optimization" } ] } ], "index": 0 }, { "bbox": [ 46, 93, 289, 237 ], "type": "text", "angle": 0, "lines": [ { "bbox": [ 46, 93, 289, 237 ], "spans": [ { "bbox": [ 46, 93, 289, 237 ], "type": "text", "content": "In order to make our skeletonization algorithms compatible with gradient-based optimization, logistic noise is added to the input during the reparametrization trick (see Section 3.3). We demonstrate the efficiency of this approach and the importance of well-tuned noise parameters, " }, { "bbox": [ 46, 93, 289, 237 ], "type": "inline_equation", "content": "\\beta" }, { "bbox": [ 46, 93, 289, 237 ], "type": "text", "content": " and " }, { "bbox": [ 46, 93, 289, 237 ], "type": "inline_equation", "content": "\\tau" }, { "bbox": [ 46, 93, 289, 237 ], "type": "text", "content": ", in a simple experiment (see Figure 6). Hereby, an input tensor is initialized with random values and passed through our skeletonization module. Using backpropagation, the tensor's values are learned so that its ultimate output resembles that of the ground truth skeleton. While several degenerate solutions that all yield the same skeleton, exist, we expect the learned tensor to ultimately resemble the skeleton itself." } ] } ], "index": 1 }, { "type": "image", "bbox": [ 55, 251, 280, 310 ], "blocks": [ { "bbox": [ 55, 251, 280, 310 ], "lines": [ { "bbox": [ 55, 251, 280, 310 ], "spans": [ { "bbox": [ 55, 251, 280, 310 ], "type": "image", "image_path": "0417adb362f41686ab1792da8cd1265a9072bd38708d15c3c55f0e0d5e40e647.jpg" } ] } ], "index": 2, "angle": 0, "type": "image_body" }, { "bbox": [ 46, 315, 288, 338 ], "lines": [ { "bbox": [ 46, 315, 288, 338 ], "spans": [ { "bbox": [ 46, 315, 288, 338 ], "type": "text", "content": "Figure 6. Experiment to test the compatibility of our skeletonization algorithm (orange box) with backpropagation." } ] } ], "index": 3, "angle": 0, "type": "image_caption" } ], "index": 2 }, { "bbox": [ 46, 351, 289, 460 ], "type": "text", "angle": 0, "lines": [ { "bbox": [ 46, 351, 289, 460 ], "spans": [ { "bbox": [ 46, 351, 289, 460 ], "type": "text", "content": "With very low levels of entropy, we observe that learning with our skeletonization module is very slow (see Figure 7). Increasing the entropy results in single passes through the skeletonization to be less faithful to the geometry and topology of the true medial axis (see Figure 8). However, averaging over repeated samples mostly recovers the true skeleton and enables learning of the correct structure. At too high entropy, convergence slows down as the obtained skeleton is not sufficiently accurate anymore." } ] } ], "index": 4 }, { "type": "image", "bbox": [ 58, 471, 274, 641 ], "blocks": [ { "bbox": [ 58, 471, 274, 641 ], "lines": [ { "bbox": [ 58, 471, 274, 641 ], "spans": [ { "bbox": [ 58, 471, 274, 641 ], "type": "image", "image_path": "31557394766c32051a9e59da63afe9db486b697e477c5e9d023956424695b98e.jpg" } ] } ], "index": 5, "angle": 0, "type": "image_body" }, { "bbox": [ 46, 643, 289, 698 ], "lines": [ { "bbox": [ 46, 643, 289, 698 ], "spans": [ { "bbox": [ 46, 643, 289, 698 ], "type": "text", "content": "Figure 7. Effect of the scale " }, { "bbox": [ 46, 643, 289, 698 ], "type": "inline_equation", "content": "\\beta" }, { "bbox": [ 46, 643, 289, 698 ], "type": "text", "content": " of the added logistic noise on the ability to propagate a gradient through our skeletonization module. Both very low entropy and very high entropy inhibit learning. Similar results can be found when varying " }, { "bbox": [ 46, 643, 289, 698 ], "type": "inline_equation", "content": "\\tau" }, { "bbox": [ 46, 643, 289, 698 ], "type": "text", "content": " (see Supplementary Material)." } ] } ], "index": 6, "angle": 0, "type": "image_caption" } ], "index": 5 }, { "type": "image", "bbox": [ 315, 72, 388, 144 ], "blocks": [ { "bbox": [ 315, 72, 388, 144 ], "lines": [ { "bbox": [ 315, 72, 388, 144 ], "spans": [ { "bbox": [ 315, 72, 388, 144 ], "type": "image", "image_path": "2bcaf50bcd236f535cb93550f886fc93d5fec07cc0bc2a369a07a3b83be03564.jpg" } ] } ], "index": 7, "angle": 0, "type": "image_body" }, { "bbox": [ 332, 145, 370, 153 ], "lines": [ { "bbox": [ 332, 145, 370, 153 ], "spans": [ { "bbox": [ 332, 145, 370, 153 ], "type": "inline_equation", "content": "\\beta = 0, n = 1" } ] } ], "index": 8, "angle": 0, "type": "image_caption" } ], "index": 7 }, { "type": "image", "bbox": [ 390, 72, 462, 144 ], "blocks": [ { "bbox": [ 390, 72, 462, 144 ], "lines": [ { "bbox": [ 390, 72, 462, 144 ], "spans": [ { "bbox": [ 390, 72, 462, 144 ], "type": "image", "image_path": "63a3f8dd631fbce31002ab009393fae5add9666fd9ae8c2cc371276f631aff70.jpg" } ] } ], "index": 9, "angle": 0, "type": "image_body" }, { "bbox": [ 405, 145, 447, 153 ], "lines": [ { "bbox": [ 405, 145, 447, 153 ], "spans": [ { "bbox": [ 405, 145, 447, 153 ], "type": "inline_equation", "content": "\\beta = 0.5, n = 1" } ] } ], "index": 10, "angle": 0, "type": "image_caption" } ], "index": 9 }, { "type": "image", "bbox": [ 465, 72, 537, 144 ], "blocks": [ { "bbox": [ 465, 72, 537, 144 ], "lines": [ { "bbox": [ 465, 72, 537, 144 ], "spans": [ { "bbox": [ 465, 72, 537, 144 ], "type": "image", "image_path": "9ffe85e5e581dddcc48674b8c2d83f3d727cac7b78efd30fb90582b37eacd835.jpg" } ] } ], "index": 11, "angle": 0, "type": "image_body" }, { "bbox": [ 480, 145, 522, 153 ], "lines": [ { "bbox": [ 480, 145, 522, 153 ], "spans": [ { "bbox": [ 480, 145, 522, 153 ], "type": "inline_equation", "content": "\\beta = 1.0, n = 1" } ] } ], "index": 12, "angle": 0, "type": "image_caption" } ], "index": 11 }, { "type": "image", "bbox": [ 314, 154, 388, 227 ], "blocks": [ { "bbox": [ 314, 154, 388, 227 ], "lines": [ { "bbox": [ 314, 154, 388, 227 ], "spans": [ { "bbox": [ 314, 154, 388, 227 ], "type": "image", "image_path": "dece5359edc5f7662db758ccecdcd342332149efb5a5ec017d010ffd0d6bb714.jpg" } ] } ], "index": 13, "angle": 0, "type": "image_body" }, { "bbox": [ 329, 228, 373, 236 ], "lines": [ { "bbox": [ 329, 228, 373, 236 ], "spans": [ { "bbox": [ 329, 228, 373, 236 ], "type": "inline_equation", "content": "\\beta = 1.0, n = 1" } ] } ], "index": 14, "angle": 0, "type": "image_caption" } ], "index": 13 }, { "type": "image", "bbox": [ 390, 154, 462, 227 ], "blocks": [ { "bbox": [ 390, 154, 462, 227 ], "lines": [ { "bbox": [ 390, 154, 462, 227 ], "spans": [ { "bbox": [ 390, 154, 462, 227 ], "type": "image", "image_path": "2803e3eb6b853b7d463c428327449b789adf9e2a23f62ca41ecfbca6e2e44561.jpg" } ] } ], "index": 15, "angle": 0, "type": "image_body" }, { "bbox": [ 402, 228, 449, 236 ], "lines": [ { "bbox": [ 402, 228, 449, 236 ], "spans": [ { "bbox": [ 402, 228, 449, 236 ], "type": "inline_equation", "content": "\\beta = 1.0, n = 10" } ] } ], "index": 16, "angle": 0, "type": "image_caption" }, { "bbox": [ 306, 241, 547, 274 ], "lines": [ { "bbox": [ 306, 241, 547, 274 ], "spans": [ { "bbox": [ 306, 241, 547, 274 ], "type": "text", "content": "Figure 8. Effect of scaling the added logistic noise " }, { "bbox": [ 306, 241, 547, 274 ], "type": "inline_equation", "content": "(\\beta)" }, { "bbox": [ 306, 241, 547, 274 ], "type": "text", "content": " in our skeletonization algorithm. Repeated sampling " }, { "bbox": [ 306, 241, 547, 274 ], "type": "inline_equation", "content": "(n)" }, { "bbox": [ 306, 241, 547, 274 ], "type": "text", "content": " mostly recovers the true skeleton." } ] } ], "index": 19, "angle": 0, "type": "image_caption" } ], "index": 15 }, { "type": "image", "bbox": [ 465, 154, 537, 227 ], "blocks": [ { "bbox": [ 465, 154, 537, 227 ], "lines": [ { "bbox": [ 465, 154, 537, 227 ], "spans": [ { "bbox": [ 465, 154, 537, 227 ], "type": "image", "image_path": "3176a7417a6314661958508c082532f0ef013479c768b932640fc21cde9daaf3.jpg" } ] } ], "index": 17, "angle": 0, "type": "image_body" }, { "bbox": [ 477, 228, 525, 236 ], "lines": [ { "bbox": [ 477, 228, 525, 236 ], "spans": [ { "bbox": [ 477, 228, 525, 236 ], "type": "inline_equation", "content": "\\beta = 1.0, n = 20" } ] } ], "index": 18, "angle": 0, "type": "image_caption" } ], "index": 17 }, { "bbox": [ 306, 293, 443, 306 ], "type": "title", "angle": 0, "lines": [ { "bbox": [ 306, 293, 443, 306 ], "spans": [ { "bbox": [ 306, 293, 443, 306 ], "type": "text", "content": "4.2. Application experiments" } ] } ], "index": 20 }, { "bbox": [ 306, 313, 519, 325 ], "type": "title", "angle": 0, "lines": [ { "bbox": [ 306, 313, 519, 325 ], "spans": [ { "bbox": [ 306, 313, 519, 325 ], "type": "text", "content": "4.2.1 Topology-aware blood vessel segmentation" } ] } ], "index": 21 }, { "bbox": [ 304, 333, 547, 453 ], "type": "text", "angle": 0, "lines": [ { "bbox": [ 304, 333, 547, 453 ], "spans": [ { "bbox": [ 304, 333, 547, 453 ], "type": "text", "content": "We explore the utility of our skeletonization methods by integrating them with a deep-learning-based segmentation algorithm for the VesSAP dataset (see Figure 9). The training of the basic U-Net incorporates the centerline Dice (clDice) loss function that encourages topology preservation across different vessel scales by comparing skeletons of the prediction and ground truth [28, 38]. The loss formulation requires a skeletonization function that is compatible with backpropagation. In all cases, we tune the weighting factor " }, { "bbox": [ 304, 333, 547, 453 ], "type": "inline_equation", "content": "\\lambda" }, { "bbox": [ 304, 333, 547, 453 ], "type": "text", "content": ", which balances the clDice loss with the standard Dice loss." } ] } ], "index": 22 }, { "type": "image", "bbox": [ 314, 464, 538, 566 ], "blocks": [ { "bbox": [ 314, 464, 538, 566 ], "lines": [ { "bbox": [ 314, 464, 538, 566 ], "spans": [ { "bbox": [ 314, 464, 538, 566 ], "type": "image", "image_path": "5b2a5a9bcb52ddd1402fc40b5b3361d1a067680bc71d64c909f9343f9666e844.jpg" } ] } ], "index": 23, "angle": 0, "type": "image_body" }, { "bbox": [ 304, 570, 547, 615 ], "lines": [ { "bbox": [ 304, 570, 547, 615 ], "spans": [ { "bbox": [ 304, 570, 547, 615 ], "type": "text", "content": "Figure 9. Deep learning pipeline for training a vessel segmentation network using the centerline Dice loss [38]. The loss formulation requires the use of a skeletonization algorithm that is compatible with backpropagation (orange box)." } ] } ], "index": 24, "angle": 0, "type": "image_caption" } ], "index": 23 }, { "bbox": [ 304, 629, 547, 715 ], "type": "text", "angle": 0, "lines": [ { "bbox": [ 304, 629, 547, 715 ], "spans": [ { "bbox": [ 304, 629, 547, 715 ], "type": "text", "content": "Using the cI Dice loss instead of a vanilla Dice loss slightly improves the topological agreement between prediction and ground truth as indicated by a lower error of the first three Betti numbers (see Table 2). Moreover, we find that using our skeletonization methods yield slightly better results than using a morphological skeletonization algorithm. Spatial accuracy, quantified by the Dice similarity coefficient (DSC)," } ] } ], "index": 25 } ], "discarded_blocks": [ { "bbox": [ 293, 748, 318, 757 ], "type": "page_number", "angle": 0, "lines": [ { "bbox": [ 293, 748, 318, 757 ], "spans": [ { "bbox": [ 293, 748, 318, 757 ], "type": "text", "content": "21400" } ] } ], "index": 26 } ], "page_size": [ 612, 792 ], "page_idx": 6 }, { "para_blocks": [ { "bbox": [ 46, 72, 288, 109 ], "type": "text", "angle": 0, "lines": [ { "bbox": [ 46, 72, 288, 109 ], "spans": [ { "bbox": [ 46, 72, 288, 109 ], "type": "text", "content": "was nearly identical in all cases. We obtain similar findings when conducting the same experiment using the DRIVE dataset (see Supplementary Material)." } ] } ], "index": 0 }, { "type": "table", "bbox": [ 54, 152, 281, 203 ], "blocks": [ { "bbox": [ 46, 120, 287, 152 ], "lines": [ { "bbox": [ 46, 120, 287, 152 ], "spans": [ { "bbox": [ 46, 120, 287, 152 ], "type": "text", "content": "Table 2. Performance of the vessel segmentation network using either a standard Dice loss ('Without') or clDice loss with different skeletonization algorithms." } ] } ], "index": 1, "angle": 0, "type": "table_caption" }, { "bbox": [ 54, 152, 281, 203 ], "lines": [ { "bbox": [ 54, 152, 281, 203 ], "spans": [ { "bbox": [ 54, 152, 281, 203 ], "type": "table", "html": "
SkeletonizationDSCβ0 errorβ1 errorβ2 error
Without0.85±0.015.1±0.83.1±0.10.8±0.3
Morphological0.85±0.014.3±0.52.8±0.20.7±0.1
Ours (Euler)0.86±0.013.5±0.22.7±0.10.4±0.1
Ours (Boolean)0.86±0.013.7±0.32.8±0.10.5±0.2
", "image_path": "a9e6a36bcc4db07bbdbfe0851d0e3de32811d4f9da2ed556549fc0a35ca5b41c.jpg" } ] } ], "index": 2, "angle": 0, "type": "table_body" } ], "index": 2 }, { "bbox": [ 46, 239, 287, 264 ], "type": "title", "angle": 0, "lines": [ { "bbox": [ 46, 239, 287, 264 ], "spans": [ { "bbox": [ 46, 239, 287, 264 ], "type": "text", "content": "4.2.2 Multimodal registration of the mandible in CT and MR images" } ] } ], "index": 3 }, { "bbox": [ 45, 272, 290, 477 ], "type": "text", "angle": 0, "lines": [ { "bbox": [ 45, 272, 290, 477 ], "spans": [ { "bbox": [ 45, 272, 290, 477 ], "type": "text", "content": "Finally, we explore whether incorporating skeletonization can improve multimodal registration of the mandible (see Figure 10). This application is motivated by the fact that bones often appear larger in MR images than in CT images. When registering segmentation masks from both modalities, the smaller mask can be orientated flexibly inside the larger one. We propose extracting the skeleton of both structures and calculating their overlap as image distance function instead. We employ a conventional registration algorithm that optimizes the image distance with respect to the rigid transformation between both images using a gradient-based optimization method, thus requiring a compatible skeletonization method. We implement this application in AirLab [33], which uses Pytorch's autograd functionality to compute the gradient based on the objective function. This allows a seamless integration of our skeletonization module in a conventional registration algorithm." } ] } ], "index": 4 }, { "type": "image", "bbox": [ 56, 490, 279, 578 ], "blocks": [ { "bbox": [ 56, 490, 279, 578 ], "lines": [ { "bbox": [ 56, 490, 279, 578 ], "spans": [ { "bbox": [ 56, 490, 279, 578 ], "type": "image", "image_path": "7cc43c54b29a0d15af49a1b280c03f559f7c4f941e882b451513091773fc4169.jpg" } ] } ], "index": 5, "angle": 0, "type": "image_body" }, { "bbox": [ 46, 582, 287, 628 ], "lines": [ { "bbox": [ 46, 582, 287, 628 ], "spans": [ { "bbox": [ 46, 582, 287, 628 ], "type": "text", "content": "Figure 10. Workflow for multimodal registration of the mandible. To compensate for the different size of the mandible in CT and MR images, the skeleton of both images are calculated (orange box) and registered instead." } ] } ], "index": 6, "angle": 0, "type": "image_caption" } ], "index": 5 }, { "bbox": [ 46, 641, 288, 715 ], "type": "text", "angle": 0, "lines": [ { "bbox": [ 46, 641, 288, 715 ], "spans": [ { "bbox": [ 46, 641, 288, 715 ], "type": "text", "content": "We report the DSC, the Hausdorff distance (HD) and the average surface distance (ASD) between the fixed and warped segmentation as proxy measure for registration accuracy (see Table 3). Our findings show registering the images based on the skeleton of their segmentation map slightly improves the alignment of both structures." } ] } ], "index": 7 }, { "type": "table", "bbox": [ 325, 112, 526, 163 ], "blocks": [ { "bbox": [ 304, 69, 547, 112 ], "lines": [ { "bbox": [ 304, 69, 547, 112 ], "spans": [ { "bbox": [ 304, 69, 547, 112 ], "type": "text", "content": "Table 3. Registration accuracy using two different loss functions: either the image distance is calculated using the full binary mask of the mandible ('Without'), or based on their skeletons obtained via one of three skeletonization algorithms." } ] } ], "index": 8, "angle": 0, "type": "table_caption" }, { "bbox": [ 325, 112, 526, 163 ], "lines": [ { "bbox": [ 325, 112, 526, 163 ], "spans": [ { "bbox": [ 325, 112, 526, 163 ], "type": "table", "html": "
SkeletonizationDSCHD [mm]ASD [mm]
Without0.38±0.0129.9±0.96.6±0.2
Morphological0.32±0.0129.2±1.16.7±0.2
Ours (Euler)0.37±0.0228.6±1.26.6±0.2
Ours (Boolean)0.37±0.0128.0±1.16.5±0.2
", "image_path": "0f63a6b80a33c1e17826d811397976dd2a08aeec463caeaa3ffc7aaac6704fcd.jpg" } ] } ], "index": 9, "angle": 0, "type": "table_body" } ], "index": 9 }, { "bbox": [ 306, 185, 454, 198 ], "type": "title", "angle": 0, "lines": [ { "bbox": [ 306, 185, 454, 198 ], "spans": [ { "bbox": [ 306, 185, 454, 198 ], "type": "text", "content": "5. Discussion and conclusion" } ] } ], "index": 10 }, { "bbox": [ 304, 207, 547, 351 ], "type": "text", "angle": 0, "lines": [ { "bbox": [ 304, 207, 547, 351 ], "spans": [ { "bbox": [ 304, 207, 547, 351 ], "type": "text", "content": "This work bridges the gap between classical skeletonization algorithms and the ability to integrate these with gradient-based optimization. We have introduced a three-dimensional skeletonization algorithm that is compatible with backpropagation, domain-agnostic and preserves an object's topology. Our method combines a characterizations of simple points, a parallelization scheme for their efficient removal, and a strategy for discretization of non-binary inputs that are all compatible with gradient-based optimization. In benchmarking experiments, we have proved the superior spatial and topological accuracy of our method compared to morphological, and neural-network-based baselines." } ] } ], "index": 11 }, { "bbox": [ 304, 353, 548, 509 ], "type": "text", "angle": 0, "lines": [ { "bbox": [ 304, 353, 548, 509 ], "spans": [ { "bbox": [ 304, 353, 548, 509 ], "type": "text", "content": "Our algorithm consists exclusively of matrix additions, multiplications, convolutional operations, activation functions and sampling from basic probability distributions. Consequently, it can be implemented in any deep learning library and can be seamlessly integrated with diverse image processing pipelines that use gradient-based optimization. We showcase this utility by applying it in two realistic medical image processing applications: semantic segmentation of blood vessels with deep learning, and automated multimodal image registration. In both cases, we find that our skeletonization algorithms allow the incorporation of topological and geometric information within the respective optimization objective, leading to modest performance gains." } ] } ], "index": 12 }, { "bbox": [ 304, 510, 548, 715 ], "type": "text", "angle": 0, "lines": [ { "bbox": [ 304, 510, 548, 715 ], "spans": [ { "bbox": [ 304, 510, 548, 715 ], "type": "text", "content": "To our knowledge, this work introduces the first topology-preserving skeletonization algorithm for gradient-based optimization. Still, we discern that there may be other, potentially more effective, approaches to create such algorithms. We hope that this work can serve as a blueprint for others to further explore skeletonization. Building upon a rich body of literature on classical skeletonization algorithms, future work could further explore alternative strategies to identify simple points [32, 35]. Similarly, past works have extensively studied schemes to efficiently remove simple points in parallel of which some may be better suited for processing on graphics processing units. Finally, the endpoint condition used during skeletonization influences the properties of the created skeleton [35, 46]. In other applications skeletal surfaces may be preferable over a medial axis or a different trade-off between a geometric and topological skeleton may be chosen. Ultimately, we envision that our method may" } ] } ], "index": 13 } ], "discarded_blocks": [ { "bbox": [ 293, 748, 317, 757 ], "type": "page_number", "angle": 0, "lines": [ { "bbox": [ 293, 748, 317, 757 ], "spans": [ { "bbox": [ 293, 748, 317, 757 ], "type": "text", "content": "21401" } ] } ], "index": 14 } ], "page_size": [ 612, 792 ], "page_idx": 7 }, { "para_blocks": [ { "bbox": [ 47, 72, 289, 109 ], "type": "text", "angle": 0, "lines": [ { "bbox": [ 47, 72, 289, 109 ], "spans": [ { "bbox": [ 47, 72, 289, 109 ], "type": "text", "content": "also be beneficial in many computer vision applications that have historically utilized skeletonization, but have since been increasingly solved using deep learning." } ] } ], "index": 0 }, { "bbox": [ 48, 121, 145, 134 ], "type": "title", "angle": 0, "lines": [ { "bbox": [ 48, 121, 145, 134 ], "spans": [ { "bbox": [ 48, 121, 145, 134 ], "type": "text", "content": "Acknowledgments" } ] } ], "index": 1 }, { "bbox": [ 47, 141, 287, 177 ], "type": "text", "angle": 0, "lines": [ { "bbox": [ 47, 141, 287, 177 ], "spans": [ { "bbox": [ 47, 141, 287, 177 ], "type": "text", "content": "We would like to thank Viktoria Thierauf for her help with the Mandible dataset. This work was funded by the Munich Center for Machine Learning." } ] } ], "index": 2 }, { "bbox": [ 48, 190, 107, 202 ], "type": "title", "angle": 0, "lines": [ { "bbox": [ 48, 190, 107, 202 ], "spans": [ { "bbox": [ 48, 190, 107, 202 ], "type": "text", "content": "References" } ] } ], "index": 3 }, { "bbox": [ 48, 209, 288, 714 ], "type": "list", "angle": 0, "index": 19, "blocks": [ { "bbox": [ 53, 209, 288, 232 ], "type": "ref_text", "angle": 0, "lines": [ { "bbox": [ 53, 209, 288, 232 ], "spans": [ { "bbox": [ 53, 209, 288, 232 ], "type": "text", "content": "[1] Dgtal: Digital geometry tools and algorithms library. http://dgtal.org.6" } ] } ], "index": 4 }, { "bbox": [ 53, 233, 288, 255 ], "type": "ref_text", "angle": 0, "lines": [ { "bbox": [ 53, 233, 288, 255 ], "spans": [ { "bbox": [ 53, 233, 288, 255 ], "type": "text", "content": "[2] Martin Abadi et al. TensorFlow: Large-scale machine learning on heterogeneous systems, 2015. 1" } ] } ], "index": 5 }, { "bbox": [ 53, 256, 287, 290 ], "type": "ref_text", "angle": 0, "lines": [ { "bbox": [ 53, 256, 287, 290 ], "spans": [ { "bbox": [ 53, 256, 287, 290 ], "type": "text", "content": "[3] Yoshua Bengio et al. Estimating or propagating gradients through stochastic neurons for conditional computation. arXiv preprint arXiv:1308.3432, 2013. 4, 5" } ] } ], "index": 6 }, { "bbox": [ 53, 291, 288, 323 ], "type": "ref_text", "angle": 0, "lines": [ { "bbox": [ 53, 291, 288, 323 ], "spans": [ { "bbox": [ 53, 291, 288, 323 ], "type": "text", "content": "[4] Gilles Bertrand. A boolean characterization of three-dimensional simple points. Pattern recognition letters, 17(2):115-124, 1996. 3, 4" } ] } ], "index": 7 }, { "bbox": [ 53, 324, 287, 357 ], "type": "ref_text", "angle": 0, "lines": [ { "bbox": [ 53, 324, 287, 357 ], "spans": [ { "bbox": [ 53, 324, 287, 357 ], "type": "text", "content": "[5] Gilles Bertrand et al. Three-dimensional thinning algorithm using subfields. In Vision Geometry III, volume 2356, pages 113-124. SPIE, 1995. 1, 4" } ] } ], "index": 8 }, { "bbox": [ 53, 358, 288, 392 ], "type": "ref_text", "angle": 0, "lines": [ { "bbox": [ 53, 358, 288, 392 ], "spans": [ { "bbox": [ 53, 358, 288, 392 ], "type": "text", "content": "[6] Gilles Bertrand et al. Powerful parallel and symmetric 3d thinning schemes based on critical kernels. Journal of Mathematical Imaging and Vision, 48:134-148, 2014. 1, 6" } ] } ], "index": 9 }, { "bbox": [ 53, 394, 288, 427 ], "type": "ref_text", "angle": 0, "lines": [ { "bbox": [ 53, 394, 288, 427 ], "spans": [ { "bbox": [ 53, 394, 288, 427 ], "type": "text", "content": "[7] Harry Blum. A transformation for extracting new descriptions of shape. Models for the perception of speech and visual form, pages 362-380, 1967. 1" } ] } ], "index": 10 }, { "bbox": [ 53, 428, 288, 450 ], "type": "ref_text", "angle": 0, "lines": [ { "bbox": [ 53, 428, 288, 450 ], "spans": [ { "bbox": [ 53, 428, 288, 450 ], "type": "text", "content": "[8] Gunilla Borgefors et al. Computing skeletons in three dimensions. Pattern recognition, 32(7):1225-1236, 1999. 1" } ] } ], "index": 11 }, { "bbox": [ 53, 452, 288, 496 ], "type": "ref_text", "angle": 0, "lines": [ { "bbox": [ 53, 452, 288, 496 ], "spans": [ { "bbox": [ 53, 452, 288, 496 ], "type": "text", "content": "[9] Ilke Demir et al. Skelneton 2019: Dataset and challenge on deep learning for geometric shape understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, 2019. 1" } ] } ], "index": 12 }, { "bbox": [ 48, 498, 288, 529 ], "type": "ref_text", "angle": 0, "lines": [ { "bbox": [ 48, 498, 288, 529 ], "spans": [ { "bbox": [ 48, 498, 288, 529 ], "type": "text", "content": "[10] Yaroslav Ganin et al. Domain-adversarial training of neural networks. The journal of machine learning research, 17(1):2096-2030, 2016. 1" } ] } ], "index": 13 }, { "bbox": [ 48, 531, 288, 576 ], "type": "ref_text", "angle": 0, "lines": [ { "bbox": [ 48, 531, 288, 576 ], "spans": [ { "bbox": [ 48, 531, 288, 576 ], "type": "text", "content": "[11] Iris AM Huijben et al. A review of the gumbel-max trick and its extensions for discrete stochasticity in machine learning. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2022. 5" } ] } ], "index": 14 }, { "bbox": [ 48, 577, 288, 599 ], "type": "ref_text", "angle": 0, "lines": [ { "bbox": [ 48, 577, 288, 599 ], "spans": [ { "bbox": [ 48, 577, 288, 599 ], "type": "text", "content": "[12] Eric Jang et al. Categorical reparameterization with gumbel-softmax. *ICLR*, 2017. 4, 5" } ] } ], "index": 15 }, { "bbox": [ 48, 600, 288, 643 ], "type": "ref_text", "angle": 0, "lines": [ { "bbox": [ 48, 600, 288, 643 ], "spans": [ { "bbox": [ 48, 600, 288, 643 ], "type": "text", "content": "[13] Koteswar Rao Jerripothula et al. Object co-skeletonization with co-segmentation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 3881-3889. IEEE, 2017. 1" } ] } ], "index": 16 }, { "bbox": [ 48, 646, 288, 678 ], "type": "ref_text", "angle": 0, "lines": [ { "bbox": [ 48, 646, 288, 678 ], "spans": [ { "bbox": [ 48, 646, 288, 678 ], "type": "text", "content": "[14] T Yung Kong et al. Digital topology: Introduction and survey. Computer Vision, Graphics, and Image Processing, 48(3):357-393, 1989. 2" } ] } ], "index": 17 }, { "bbox": [ 48, 679, 288, 714 ], "type": "ref_text", "angle": 0, "lines": [ { "bbox": [ 48, 679, 288, 714 ], "spans": [ { "bbox": [ 48, 679, 288, 714 ], "type": "text", "content": "[15] Ta-Chih Lee et al. Building skeleton models via 3-D medial surface axis thinning algorithms. CVGIP: Graphical Models and Image Processing, 56(6):462-478, 1994. 1, 2, 3, 4" } ] } ], "index": 18 } ], "sub_type": "ref_text" }, { "bbox": [ 307, 73, 547, 713 ], "type": "list", "angle": 0, "index": 36, "blocks": [ { "bbox": [ 307, 73, 547, 106 ], "type": "ref_text", "angle": 0, "lines": [ { "bbox": [ 307, 73, 547, 106 ], "spans": [ { "bbox": [ 307, 73, 547, 106 ], "type": "text", "content": "[16] Jiang-Jiang Liu et al. Dynamic feature integration for simultaneous detection of salient object, edge, and skeleton. IEEE Transactions on Image Processing, 29:8652-8667, 2020. 1" } ] } ], "index": 20 }, { "bbox": [ 307, 108, 547, 140 ], "type": "ref_text", "angle": 0, "lines": [ { "bbox": [ 307, 108, 547, 140 ], "spans": [ { "bbox": [ 307, 108, 547, 140 ], "type": "text", "content": "[17] Steven Lobregt et al. Three-dimensional skeletonization: principle and algorithm. IEEE Transactions on pattern analysis and machine intelligence, 2(1):75-77, 1980. 3" } ] } ], "index": 21 }, { "bbox": [ 307, 142, 547, 175 ], "type": "ref_text", "angle": 0, "lines": [ { "bbox": [ 307, 142, 547, 175 ], "spans": [ { "bbox": [ 307, 142, 547, 175 ], "type": "text", "content": "[18] C Min Ma et al. A fully parallel 3D thinning algorithm and its applications. Computer vision and image understanding, 64(3):420-433, 1996. 1, 4" } ] } ], "index": 22 }, { "bbox": [ 307, 177, 547, 231 ], "type": "ref_text", "angle": 0, "lines": [ { "bbox": [ 307, 177, 547, 231 ], "spans": [ { "bbox": [ 307, 177, 547, 231 ], "type": "text", "content": "[19] Chris J. Maddison et al. The concrete distribution: A continuous relaxation of discrete random variables. In Proceedings of the International Conference on Learning Representations, International Conference on Learning Representations, 2017. 4, 5" } ] } ], "index": 23 }, { "bbox": [ 307, 233, 547, 266 ], "type": "ref_text", "angle": 0, "lines": [ { "bbox": [ 307, 233, 547, 266 ], "spans": [ { "bbox": [ 307, 233, 547, 266 ], "type": "text", "content": "[20] Petros Maragos et al. Morphological skeleton representation and coding of binary images. IEEE Transactions on Acoustics, Speech, and Signal Processing, 34(5):1228-1244, 1986. 1" } ] } ], "index": 24 }, { "bbox": [ 307, 268, 547, 300 ], "type": "ref_text", "angle": 0, "lines": [ { "bbox": [ 307, 268, 547, 300 ], "spans": [ { "bbox": [ 307, 268, 547, 300 ], "type": "text", "content": "[21] David G Morgenthaler. Three-dimensional simple points: serial erosion, parallel thinning and skeletonization. TR-1005, 1981. 2, 3, 4" } ] } ], "index": 25 }, { "bbox": [ 307, 302, 547, 346 ], "type": "ref_text", "angle": 0, "lines": [ { "bbox": [ 307, 302, 547, 346 ], "spans": [ { "bbox": [ 307, 302, 547, 346 ], "type": "text", "content": "[22] Bryan S Morse et al. Multiscale medial analysis of medical images. In Biennial International Conference on Information Processing in Medical Imaging, pages 112-131. Springer, 1993. 1" } ] } ], "index": 26 }, { "bbox": [ 307, 348, 547, 391 ], "type": "ref_text", "angle": 0, "lines": [ { "bbox": [ 307, 348, 547, 391 ], "spans": [ { "bbox": [ 307, 348, 547, 391 ], "type": "text", "content": "[23] Sabari Nathan et al. Skeletonnet: Shape pixel to skeleton pixel. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, pages 0–0, 2019. 1" } ] } ], "index": 27 }, { "bbox": [ 307, 393, 547, 437 ], "type": "ref_text", "angle": 0, "lines": [ { "bbox": [ 307, 393, 547, 437 ], "spans": [ { "bbox": [ 307, 393, 547, 437 ], "type": "text", "content": "[24] Nam Hoang Nguyen. U-net based skeletonization and bag of tricks. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2105-2109, 2021. 1, 6" } ] } ], "index": 28 }, { "bbox": [ 307, 439, 547, 483 ], "type": "ref_text", "angle": 0, "lines": [ { "bbox": [ 307, 439, 547, 483 ], "spans": [ { "bbox": [ 307, 439, 547, 483 ], "type": "text", "content": "[25] Kálmán Palágyi et al. A sequential 3D thinning algorithm and its medical applications. In Biennial International Conference on Information Processing in Medical Imaging, pages 409-415. Springer, 2001. 1" } ] } ], "index": 29 }, { "bbox": [ 307, 484, 547, 529 ], "type": "ref_text", "angle": 0, "lines": [ { "bbox": [ 307, 484, 547, 529 ], "spans": [ { "bbox": [ 307, 484, 547, 529 ], "type": "text", "content": "[26] Oleg Panichev et al. U-net based convolutional neural network for skeleton extraction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, 2019. 1, 6" } ] } ], "index": 30 }, { "bbox": [ 307, 531, 547, 553 ], "type": "ref_text", "angle": 0, "lines": [ { "bbox": [ 307, 531, 547, 553 ], "spans": [ { "bbox": [ 307, 531, 547, 553 ], "type": "text", "content": "[27] Adam Paszke et al. Automatic differentiation in pytorch. NIPS 2017 Workshop Autodiff, 2017. 1, 5" } ] } ], "index": 31 }, { "bbox": [ 307, 555, 547, 620 ], "type": "ref_text", "angle": 0, "lines": [ { "bbox": [ 307, 555, 547, 620 ], "spans": [ { "bbox": [ 307, 555, 547, 620 ], "type": "text", "content": "[28] Olaf Ronneberger et al. U-net: Convolutional networks for biomedical image segmentation. In Medical Image Computing and Computer-Assisted Intervention-MICCAI 2015: 18th International Conference, Munich, Germany, October 5-9, 2015, Proceedings, Part III 18, pages 234-241. Springer, 2015. 7" } ] } ], "index": 32 }, { "bbox": [ 307, 622, 547, 654 ], "type": "ref_text", "angle": 0, "lines": [ { "bbox": [ 307, 622, 547, 654 ], "spans": [ { "bbox": [ 307, 622, 547, 654 ], "type": "text", "content": "[29] Frank Rosenblatt. The perceptron, a perceiving and recognizing automaton. Cornell Aeronautical Laboratory, 1957. 5" } ] } ], "index": 33 }, { "bbox": [ 307, 656, 547, 679 ], "type": "ref_text", "angle": 0, "lines": [ { "bbox": [ 307, 656, 547, 679 ], "spans": [ { "bbox": [ 307, 656, 547, 679 ], "type": "text", "content": "[30] Azriel Rosenfeld et al. Digital picture processing. Academic press, 1976. 2" } ] } ], "index": 34 }, { "bbox": [ 307, 680, 547, 713 ], "type": "ref_text", "angle": 0, "lines": [ { "bbox": [ 307, 680, 547, 713 ], "spans": [ { "bbox": [ 307, 680, 547, 713 ], "type": "text", "content": "[31] Punam K Saha et al. A new shape preserving parallel thinning algorithm for 3D digital images. Pattern recognition, 30(12):1939-1955, 1997. 1, 4" } ] } ], "index": 35 } ], "sub_type": "ref_text" } ], "discarded_blocks": [ { "bbox": [ 293, 748, 318, 757 ], "type": "page_number", "angle": 0, "lines": [ { "bbox": [ 293, 748, 318, 757 ], "spans": [ { "bbox": [ 293, 748, 318, 757 ], "type": "text", "content": "21402" } ] } ], "index": 37 } ], "page_size": [ 612, 792 ], "page_idx": 8 }, { "para_blocks": [ { "bbox": [ 48, 73, 287, 579 ], "type": "list", "angle": 0, "index": 15, "blocks": [ { "bbox": [ 48, 73, 287, 105 ], "type": "ref_text", "angle": 0, "lines": [ { "bbox": [ 48, 73, 287, 105 ], "spans": [ { "bbox": [ 48, 73, 287, 105 ], "type": "text", "content": "[32] Punam K Saha et al. A survey on skeletonization algorithms and their applications. Pattern recognition letters, 76:3-12, 2016. 1, 3, 8" } ] } ], "index": 0 }, { "bbox": [ 49, 107, 287, 129 ], "type": "ref_text", "angle": 0, "lines": [ { "bbox": [ 49, 107, 287, 129 ], "spans": [ { "bbox": [ 49, 107, 287, 129 ], "type": "text", "content": "[33] Robin Sandkühler et al. Airlab: autograd image registration laboratory. arXiv preprint arXiv:1806.09907, 2018. 8" } ] } ], "index": 1 }, { "bbox": [ 49, 130, 287, 152 ], "type": "ref_text", "angle": 0, "lines": [ { "bbox": [ 49, 130, 287, 152 ], "spans": [ { "bbox": [ 49, 130, 287, 152 ], "type": "text", "content": "[34] Doron Shaked et al. Pruning medial axes. Computer vision and image understanding, 69(2):156-169, 1998. 4" } ] } ], "index": 2 }, { "bbox": [ 49, 153, 287, 185 ], "type": "ref_text", "angle": 0, "lines": [ { "bbox": [ 49, 153, 287, 185 ], "spans": [ { "bbox": [ 49, 153, 287, 185 ], "type": "text", "content": "[35] Wei Shen et al. Skeleton pruning as trade-off between skeleton simplicity and reconstruction error. Science China Information Sciences, 56:1-14, 2013. 8" } ] } ], "index": 3 }, { "bbox": [ 49, 187, 287, 230 ], "type": "ref_text", "angle": 0, "lines": [ { "bbox": [ 49, 187, 287, 230 ], "spans": [ { "bbox": [ 49, 187, 287, 230 ], "type": "text", "content": "[36] Wei Shen et al. Object skeleton extraction in natural images by fusing scale-associated deep side outputs. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 222-230, 2016. 1" } ] } ], "index": 4 }, { "bbox": [ 49, 232, 287, 274 ], "type": "ref_text", "angle": 0, "lines": [ { "bbox": [ 49, 232, 287, 274 ], "spans": [ { "bbox": [ 49, 232, 287, 274 ], "type": "text", "content": "[37] Wei Shen et al. Deepskeleton: Learning multi-task scale-associated deep side outputs for object skeleton extraction in natural images. IEEE Transactions on Image Processing, 26(11):5298-5311, 2017. 1" } ] } ], "index": 5 }, { "bbox": [ 49, 277, 287, 319 ], "type": "ref_text", "angle": 0, "lines": [ { "bbox": [ 49, 277, 287, 319 ], "spans": [ { "bbox": [ 49, 277, 287, 319 ], "type": "text", "content": "[38] Suprosanna Shit et al. cldice-a novel topology-preserving loss function for tubular structure segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16560-16569, 2021. 1, 6, 7" } ] } ], "index": 6 }, { "bbox": [ 49, 322, 287, 353 ], "type": "ref_text", "angle": 0, "lines": [ { "bbox": [ 49, 322, 287, 353 ], "spans": [ { "bbox": [ 49, 322, 287, 353 ], "type": "text", "content": "[39] Joes Staal et al. Ridge-based vessel segmentation in color images of the retina. IEEE transactions on medical imaging, 23(4):501-509, 2004. 5" } ] } ], "index": 7 }, { "bbox": [ 49, 356, 287, 376 ], "type": "ref_text", "angle": 0, "lines": [ { "bbox": [ 49, 356, 287, 376 ], "spans": [ { "bbox": [ 49, 356, 287, 376 ], "type": "text", "content": "[40] David Thibault et al. Terrain reconstruction from contours by skeleton construction. *GeoInformatica*, 4:349–373, 2000. 1" } ] } ], "index": 8 }, { "bbox": [ 49, 378, 287, 409 ], "type": "ref_text", "angle": 0, "lines": [ { "bbox": [ 49, 378, 287, 409 ], "spans": [ { "bbox": [ 49, 378, 287, 409 ], "type": "text", "content": "[41] Mihail Ivilinov Todorov et al. Machine learning analysis of whole mouse brain vasculature. Nature methods, 17(4):442-449, 2020. 5" } ] } ], "index": 9 }, { "bbox": [ 49, 412, 287, 443 ], "type": "ref_text", "angle": 0, "lines": [ { "bbox": [ 49, 412, 287, 443 ], "spans": [ { "bbox": [ 49, 412, 287, 443 ], "type": "text", "content": "[42] YF Tsao et al. A parallel thinning algorithm for 3-D pictures. Computer graphics and image processing, 17(4):315-331, 1981. 1, 4" } ] } ], "index": 10 }, { "bbox": [ 49, 446, 287, 478 ], "type": "ref_text", "angle": 0, "lines": [ { "bbox": [ 49, 446, 287, 478 ], "spans": [ { "bbox": [ 49, 446, 287, 478 ], "type": "text", "content": "[43] David Vazquez et al. Virtual and real world adaptation for pedestrian detection. IEEE transactions on pattern analysis and machine intelligence, 36(4):797-809, 2013. 1" } ] } ], "index": 11 }, { "bbox": [ 49, 480, 287, 511 ], "type": "ref_text", "angle": 0, "lines": [ { "bbox": [ 49, 480, 287, 511 ], "spans": [ { "bbox": [ 49, 480, 287, 511 ], "type": "text", "content": "[44] Mario Viti et al. Coronary artery centerline tracking with the morphological skeleton loss. In Proc. ICIP, pages 2741-2745, 2022. 1, 6" } ] } ], "index": 12 }, { "bbox": [ 49, 514, 287, 545 ], "type": "ref_text", "angle": 0, "lines": [ { "bbox": [ 49, 514, 287, 545 ], "spans": [ { "bbox": [ 49, 514, 287, 545 ], "type": "text", "content": "[45] Feng Zhao et al. Preprocessing and postprocessing for skeleton-based fingerprint minutiae extraction. Pattern Recognition, 40(4):1270-1281, 2007. 1" } ] } ], "index": 13 }, { "bbox": [ 49, 548, 287, 579 ], "type": "ref_text", "angle": 0, "lines": [ { "bbox": [ 49, 548, 287, 579 ], "spans": [ { "bbox": [ 49, 548, 287, 579 ], "type": "text", "content": "[46] Yong Zhou et al. Efficient skeletonization of volumetric objects. IEEE Transactions on visualization and computer graphics, 5(3):196-209, 1999. 1, 4, 8" } ] } ], "index": 14 } ], "sub_type": "ref_text" } ], "discarded_blocks": [ { "bbox": [ 293, 749, 317, 757 ], "type": "page_number", "angle": 0, "lines": [ { "bbox": [ 293, 749, 317, 757 ], "spans": [ { "bbox": [ 293, 749, 317, 757 ], "type": "text", "content": "21403" } ] } ], "index": 16 } ], "page_size": [ 612, 792 ], "page_idx": 9 } ], "_backend": "vlm", "_version_name": "2.6.4" }