[
  {
    "chunk_id": "ee05bb82-19af-40f6-ae4c-f9f64cec1ca0",
    "text": "Center for Research in Computer Vision (CRCV),\nSchool of Computer Science, University of Central Florida, Orlando, FL. The state of the art lung nodule detection studies rely on2018\ncomputationally expensive multi-stage frameworks to detect nodules from\nCT scans. To address this computational challenge and provide better\nperformance, in this paper we propose S4ND, a new deep learning basedJun\nmethod for lung nodule detection. Our approach uses a single feed for-\n3 ward pass of a single network for detection and provides better performance when compared to the current literature. The whole detection\npipeline is designed as a single 3D Convolutional Neural Network (CNN)\nwith dense connections, trained in an end-to-end manner. S4ND does\nnot require any further post-processing or user guidance to refine detection results.",
    "paper_id": "1805.02279",
    "title": "S4ND: Single-Shot Single-Scale Lung Nodule Detection",
    "authors": [
      "Naji Khosravan",
      "Ulas Bagci"
    ],
    "published_date": "2018-05-06",
    "primary_category": "cs.CV",
    "arxiv_url": "http://arxiv.org/abs/1805.02279v2",
    "chunk_index": 1,
    "total_chunks": 20,
    "char_count": 829,
    "word_count": 125,
    "chunking_strategy": "semantic"
  },
  {
    "chunk_id": "d163a366-e38d-423c-8d06-b14d718ce4a0",
    "text": "Experimentally, we compared our network with the current\nstate-of-the-art object detection network (SSD) in computer vision as[cs.CV]\nwell as the state-of-the-art published method for lung nodule detection\n(3D DCNN). We used publically available 888 CT scans from LUNA\nchallenge dataset and showed that the proposed method outperforms the\ncurrent literature both in terms of efficiency and accuracy by achieving\nan average FROC-score of 0.897. We also provide an in-depth analysis\nof our proposed network to shed light on the unclear paradigms of tiny\nobject detection. Keywords: Detection · Single-shot · CNN · Lung Nodule · Dense CNN\n· Tiny Object Detection. of imaging data (CT scans) every day. Computer Aided Detection (CAD) systems are designed to help radiologists in the screening process. However, automatic detection of lung nodules with CADs remains a challenging task. One\nreason is the high variation in texture, shape, and position of nodules in CT\nscans, and their similarity with other nearby structures. Another reason is the\ndiscrepancy between the large search space (i.e., entire lung fields) and respectively tiny nature of the nodules. Detection of tiny/small objects has remained a\nvery challenging task in computer vision, which so far has only been solved using\ncomputationally expensive multi-stage frameworks.",
    "paper_id": "1805.02279",
    "title": "S4ND: Single-Shot Single-Scale Lung Nodule Detection",
    "authors": [
      "Naji Khosravan",
      "Ulas Bagci"
    ],
    "published_date": "2018-05-06",
    "primary_category": "cs.CV",
    "arxiv_url": "http://arxiv.org/abs/1805.02279v2",
    "chunk_index": 2,
    "total_chunks": 20,
    "char_count": 1336,
    "word_count": 203,
    "chunking_strategy": "semantic"
  },
  {
    "chunk_id": "75d381c3-dcd7-4288-a098-dcdaa656dd4c",
    "text": "Current sate of art methods for lung nodule detection follow the same multi-stage detection frameworks as\nin other computer vision areas. The literature for lung nodule detection and diagnosis is vast. To date, the\ncommon strategy for all available CAD systems for lung nodule detection is to\nuse a candidate identification step (also known as region proposal). While some\nof these studies apply low-level appearance based features as a prior to drive\nthis identification task [8], others use shape and size information [5]. Related to\ndeep learning based methods, Ypsilantis et al. proposed to use recurrent neural\nnetworks in a patch based strategy to improve nodule detection [11]. Krishnamurthy et al. proposed to detect candidates using a 2D multi-step segmentation\nprocess. Then a group of hand-crafted features were extracted, followed by a\ntwo-stage classification of candidates [5]. In a similar fashion, Huang et al. proposed a geometric model based candidate detection method which followed by\na 3D CNN to reduce number of FPs [4]. Golan et al. used a deep 3D CNN\nwith a small input patch of 5 × 20 × 20 for lung nodule detection. The network\nwas applied to the lung CT volume multiple times using a sliding window and\nexhaustive search strategy to output a probability map over the volume [3]. There has, also, been detailed investigations of high-level discriminatory information extraction using deep networks to perform a better FP reduction [10]. Setio et al. used 9 separate 2D convolutional neural networks trained on 9 different views of candidates, followed by a fusion strategy to perform FP reduction [10]. Another study used a modified version of Faster R-CNN, state of the\nart object detector at the time, for candidate detection and a patch based 3D\nCNN for FP reduction step [1]. However, all these methods are computationally\ninefficient (e.g., exhaustive use of sliding windows over feature maps), and often\ncomputed in 2D manner, not appreciating the 3D nature of the nodule space.",
    "paper_id": "1805.02279",
    "title": "S4ND: Single-Shot Single-Scale Lung Nodule Detection",
    "authors": [
      "Naji Khosravan",
      "Ulas Bagci"
    ],
    "published_date": "2018-05-06",
    "primary_category": "cs.CV",
    "arxiv_url": "http://arxiv.org/abs/1805.02279v2",
    "chunk_index": 3,
    "total_chunks": 20,
    "char_count": 2010,
    "word_count": 326,
    "chunking_strategy": "semantic"
  },
  {
    "chunk_id": "9a0eb2d2-a96c-4e11-8af2-04ec4f65ad44",
    "text": "It\nis worth mentioning that patch based methods are 3D but they suffer from the\nsame computational burdens, as well as missing the entire notion of 3D nodule\nspace due to limited information available in the patches. Our Contributions: We resolve the aforementioned issues by proposing\na completely 3D deep network architecture designed to detect lung nodules\nin a single shot using a single-scale network. To the best of our knowledge,\nthis is the first study to perform lung nodule detection in one step. Specific\nto the architecture design of the deep network, we make use of convolution\nblocks with dense connections for this problem, making one step nodule detection\ncomputationally feasible. We also investigate and justify the effect of different\ndown-sampling methods in our network due to its important role for tiny object\ndetection. Lastly, we argue that lung nodule detection, as opposed to object\ndetection in natural images, can be done with high accuracy using only a single\nscale network when network is carefully designed with its hyper-parameters. Fig. 1 shows the overview of the proposed method for lung nodule detection in\na single shot. The input to our network is a 3D volume of a lung CT scan. proposed 3D densely connected Convolutional Neural Network (CNN) divides\nthe input volume into a grid of size S × S × T cells. We model lung nodule detection as a cell-wise classification problem, done simultaneously for all the cells. Unlike commonly used region proposal networks, our proposed network is able\nto reason the presence of nodule in a cell using global contextual information,\nbased on the whole 3D input volume. Dense Block Dense Block Our framework, named S4ND, models nodule detection as a cell-wise classification of the input volume. The input volume is divided by a 16 × 16 × 8 grid and\nis passed through a newly designed 3D dense CNN. The output is a probability map\nindicating the presence of a nodule in each cell.",
    "paper_id": "1805.02279",
    "title": "S4ND: Single-Shot Single-Scale Lung Nodule Detection",
    "authors": [
      "Naji Khosravan",
      "Ulas Bagci"
    ],
    "published_date": "2018-05-06",
    "primary_category": "cs.CV",
    "arxiv_url": "http://arxiv.org/abs/1805.02279v2",
    "chunk_index": 4,
    "total_chunks": 20,
    "char_count": 1956,
    "word_count": 328,
    "chunking_strategy": "semantic"
  },
  {
    "chunk_id": "f6c15a83-c487-46c7-b92d-d888298e5cf1",
    "text": "2.1 Single-Scale Detection As opposed to object detection in natural scenes, we show that lung nodule\ndetection can be performed efficiently and with high accuracy in a single scale. Current literature reports the most frequently observed nodule sizes fall within\n3mms to 32mms [9], most of which are less than 9mm and are considered as small\n(def. American Thoracic Society).",
    "paper_id": "1805.02279",
    "title": "S4ND: Single-Shot Single-Scale Lung Nodule Detection",
    "authors": [
      "Naji Khosravan",
      "Ulas Bagci"
    ],
    "published_date": "2018-05-06",
    "primary_category": "cs.CV",
    "arxiv_url": "http://arxiv.org/abs/1805.02279v2",
    "chunk_index": 5,
    "total_chunks": 20,
    "char_count": 376,
    "word_count": 60,
    "chunking_strategy": "semantic"
  },
  {
    "chunk_id": "dd774c58-15b2-4e89-a720-63050cf2e572",
    "text": "Nodules less than 3mm in size are the most\ndifficult to detect due to their tiny nature and high similarities to vessels. Based\non the statistics of nodule size and the evidence in literature, we hypothesize\nthat a single scale framework with the grid size that we defined (16 × 16 × 8\nleading to the cell sized of 32 × 32 × 8 on a volume of size 512 × 512 × 8) is\nsufficient to fit all the expected nodule sizes and provide good detection results\nwithout the need to increase the algorithmic complexity to multi-scale. This has\nbeen partially proven in other multi-scale studies [2]. 2.2 Dense and Deeper Convolution Blocks Improve Detection The loss of low-level information throughout a network causes either a high\nnumber of false positives or low sensitivity. One efficient way that helps the\nflow of information in a network and keeps this low-level information, combining\nit with the high level information, is the use of dense connections inside the\nconvolution blocks. We empirically show that deeper densely-connected blocks\nprovide better detection results. This, however, comes with the cost of more\ncomputation. In our experiments we found that dense blocks with 6 convolution\nlayers provide a good balance of detection accuracy and computational efficiency. 2.3 Max-Pooling Improves Detection As we go deeper in a CNN, it is desired to pick the most descriptive features and\npass only those to the next layers. Recently, architectures for object detection in natural images preferred the use of convolutions with stride 2 instead of\npooling [7]. In the context of tiny object detection, this feature reduction plays\nan important role. Since our objects of interest are small, if we carelessly pick\nthe features to propagate we can easily lose the objects of interest through the\nnetwork and end up with a sub-optimal model.",
    "paper_id": "1805.02279",
    "title": "S4ND: Single-Shot Single-Scale Lung Nodule Detection",
    "authors": [
      "Naji Khosravan",
      "Ulas Bagci"
    ],
    "published_date": "2018-05-06",
    "primary_category": "cs.CV",
    "arxiv_url": "http://arxiv.org/abs/1805.02279v2",
    "chunk_index": 6,
    "total_chunks": 20,
    "char_count": 1837,
    "word_count": 306,
    "chunking_strategy": "semantic"
  },
  {
    "chunk_id": "f238feec-83e6-45c1-a9a5-b0cdcd9a7ff3",
    "text": "In theory, the goal is to have\nas less pooling as possible. Also, it is desired to have this feature sampling step\nin a way that information loss is minimized. There are multiple approaches for\nsampling information through the network. Average pooling, max pooling and\nconvolutions with stride 2 are some of the options. In our experiments, we showed\nthat max pooling is the best choice of feature sampling for our task as it selects\nthe most discriminative feature in the network. Also, we showed that convolution\nlayers with stride of 2 are performing better compared to average pooling. The\nreason is that convolution with stride 2 is very similar in its nature to weighted\naveraging with the weights being learned in a data driven manner. 2.4 Proposed 3D Deep Network Architecture Our network architecture consists of 36, 3D convolution layers, 4 max-pooling\nlayers and a sigmoid activation function at the end. 30 of convolution layers\nform 5 blocks with dense connections and without pooling, which enhance lowlevel information along with high-level information, and the remainder form the\ntransition layers.",
    "paper_id": "1805.02279",
    "title": "S4ND: Single-Shot Single-Scale Lung Nodule Detection",
    "authors": [
      "Naji Khosravan",
      "Ulas Bagci"
    ],
    "published_date": "2018-05-06",
    "primary_category": "cs.CV",
    "arxiv_url": "http://arxiv.org/abs/1805.02279v2",
    "chunk_index": 7,
    "total_chunks": 20,
    "char_count": 1114,
    "word_count": 180,
    "chunking_strategy": "semantic"
  },
  {
    "chunk_id": "eb9ad5fd-b545-494f-ae95-a621b6b88fc6",
    "text": "The details of our architecture can be seen in Fig. 2. The input\nto our network is 512 × 512 × 8 and the output is a 16 × 16 × 8 probability map. Each cell in the output corresponds to a cell of the original image divided by a\n16 × 16 × 8 grid and decides whether there is a nodule in that cell or not.",
    "paper_id": "1805.02279",
    "title": "S4ND: Single-Shot Single-Scale Lung Nodule Detection",
    "authors": [
      "Naji Khosravan",
      "Ulas Bagci"
    ],
    "published_date": "2018-05-06",
    "primary_category": "cs.CV",
    "arxiv_url": "http://arxiv.org/abs/1805.02279v2",
    "chunk_index": 8,
    "total_chunks": 20,
    "char_count": 302,
    "word_count": 68,
    "chunking_strategy": "semantic"
  },
  {
    "chunk_id": "72b9c925-f8ef-4d07-830e-bf1c3873d15d",
    "text": "Dense Block\n512 (Growth rate: 16\nDense Block Num of convs: 6)\nDense Block Dense Block (Growth rate: 16 Dense Block rate: 16 (Growth rate: Num of convs: 6) (Growth 32 (Growth rate: Num of convs: 64 16 Num of convs:\nNum of convs: 6) 6) 8 8\nDense Block: Conv kernel sizes are 3x3x3 Transition layer: 1x1 conv with output channel of 4xGrowth Rate\nConv: Conv layer with kernel size 1x1x1 Maxpool Sigmoid Input to the network is a 512 × 512 × 8 volume and output is a 16 × 16 × 8\nprobability map representing likelihood of nodule presence. Our network has 5 dense\nblocks each having 6 conv. layers. The growth rates of blocks 1 to 5 is 16, 16, 16, 32, 64\nrespectively.",
    "paper_id": "1805.02279",
    "title": "S4ND: Single-Shot Single-Scale Lung Nodule Detection",
    "authors": [
      "Naji Khosravan",
      "Ulas Bagci"
    ],
    "published_date": "2018-05-06",
    "primary_category": "cs.CV",
    "arxiv_url": "http://arxiv.org/abs/1805.02279v2",
    "chunk_index": 9,
    "total_chunks": 20,
    "char_count": 662,
    "word_count": 128,
    "chunking_strategy": "semantic"
  },
  {
    "chunk_id": "e35a5f71-0c0c-468e-9d26-f25ccff8b25c",
    "text": "The network has 4 transition layers and 4 max-pooling layers. The last\nblock is followed by a convolution layer with kernel size 1 × 1 × 1 and output channel\nof 1 and a sigmoid activation function. Densely connected convolution blocks: As stated, our network consists\nof 5 densely connected blocks, each block containing 6 convolution layers with",
    "paper_id": "1805.02279",
    "title": "S4ND: Single-Shot Single-Scale Lung Nodule Detection",
    "authors": [
      "Naji Khosravan",
      "Ulas Bagci"
    ],
    "published_date": "2018-05-06",
    "primary_category": "cs.CV",
    "arxiv_url": "http://arxiv.org/abs/1805.02279v2",
    "chunk_index": 10,
    "total_chunks": 20,
    "char_count": 346,
    "word_count": 58,
    "chunking_strategy": "semantic"
  },
  {
    "chunk_id": "02d14b0b-75d9-46bb-a748-39367096bf39",
    "text": "an output channel of g, which is the growth rate of that block. Inside the blocks,\neach layer receives all the preceding layers' feature maps as inputs. Fig. 2 (top\nright) illustrates the layout of a typical dense block. Dense connections help the\nflow of information inside the network.",
    "paper_id": "1805.02279",
    "title": "S4ND: Single-Shot Single-Scale Lung Nodule Detection",
    "authors": [
      "Naji Khosravan",
      "Ulas Bagci"
    ],
    "published_date": "2018-05-06",
    "primary_category": "cs.CV",
    "arxiv_url": "http://arxiv.org/abs/1805.02279v2",
    "chunk_index": 11,
    "total_chunks": 20,
    "char_count": 287,
    "word_count": 49,
    "chunking_strategy": "semantic"
  },
  {
    "chunk_id": "17fa6ea3-dfcf-4e4c-9ee1-66f5f96a441e",
    "text": "Assume x0 is the input volume to the\nblock and xi is the output feature map of layer i inside the block. Each layer\nis a non-linear function Fi, which in our case is a composition of convolution,\nbatch normalization (BN) and rectifier linear unit (ReLU). With dense connections, each layer receives a concatenation of all previous layers' feature maps as\ninput xi = Fi([x0, x1, ..., xi−1]), where xi is the output feature map from layer i\nand [x0, x1, ..., xi−1] is the channel-wise concatenation of previous layers' feature\nmaps. Growth rate (GR): is the number of feature maps that each layer Fi\nproduces in the block. This number is fixed for each block but it can change\nfrom one block to the other. Assume the number of channels in the input layer\nof a block is c0 and the block has i convolution layers with a growth rate of g. Then the output of the block will have c0 + (i −1)g channels. Transition layers: as can be seen in the above formulations, the number of\nfeature maps inside each dense block increases dramatically. Transition layers are\n1×1×1 convolution layers with 4×g output channels, where g is the growth rate\nof previous block. Using a convolution with kernel size of 1×1×1 compresses the\ninformation channel-wise and reduces the total number of channels throughout\nthe network. Training the network: The created ground truths for training our network\nare 3D volumes with size 16 × 16 × 8. Each element in this volume corresponds\nto a cell in the input image and has label 1 if a nodule exists in that cell and\n0 otherwise.",
    "paper_id": "1805.02279",
    "title": "S4ND: Single-Shot Single-Scale Lung Nodule Detection",
    "authors": [
      "Naji Khosravan",
      "Ulas Bagci"
    ],
    "published_date": "2018-05-06",
    "primary_category": "cs.CV",
    "arxiv_url": "http://arxiv.org/abs/1805.02279v2",
    "chunk_index": 12,
    "total_chunks": 20,
    "char_count": 1546,
    "word_count": 276,
    "chunking_strategy": "semantic"
  },
  {
    "chunk_id": "83862a61-5bbf-44fb-a57a-b35463b7c11b",
    "text": "The design of our network allows for an end-to-end training. We\nmodel detection as a cell wise classification of input which is done in one feed\nforward path of the network in one shot. This formulation detects all the nodules\nin the given volume simultaneously. The loss function for training our network\nis weighted cross-entropy defined as: kn\nL(Y (n), f(X(n)) = X −yi log(f(xi)), (1)\ni=1\nwhere Y s are the labels and Xs are the inputs.",
    "paper_id": "1805.02279",
    "title": "S4ND: Single-Shot Single-Scale Lung Nodule Detection",
    "authors": [
      "Naji Khosravan",
      "Ulas Bagci"
    ],
    "published_date": "2018-05-06",
    "primary_category": "cs.CV",
    "arxiv_url": "http://arxiv.org/abs/1805.02279v2",
    "chunk_index": 13,
    "total_chunks": 20,
    "char_count": 439,
    "word_count": 78,
    "chunking_strategy": "semantic"
  },
  {
    "chunk_id": "96245939-6d8c-44f0-83ba-3acaa2fbe2fb",
    "text": "3 Experiments and Results Data and evaluation: To evaluate detection performance of S4ND, we used\nLung Nodule Analysis (LUNA16) Challenge dataset (consisting of a total of 888\nchest CT scans, slice thickness< 2.5 mm, with ground truth nodule locations). For the training, we performed a simple data augmentation by shifting the images in 4 directions by 32 pixels. We sampled the 3D volumes for training so\nthat nodules appear in random locations to avoid bias toward location of nodules.",
    "paper_id": "1805.02279",
    "title": "S4ND: Single-Shot Single-Scale Lung Nodule Detection",
    "authors": [
      "Naji Khosravan",
      "Ulas Bagci"
    ],
    "published_date": "2018-05-06",
    "primary_category": "cs.CV",
    "arxiv_url": "http://arxiv.org/abs/1805.02279v2",
    "chunk_index": 14,
    "total_chunks": 20,
    "char_count": 488,
    "word_count": 79,
    "chunking_strategy": "semantic"
  },
  {
    "chunk_id": "aa400315-a1e3-450f-80b0-58f930134a04",
    "text": "We performed 10-fold cross validation to evaluate our method by following\nthe LUNA challenge guidelines. Free-Response Receiver Operating Characteristic (FROC) analysis has been conducted to calculate sensitivity and specificity [6]. Suggested by the challenge organizers, sensitivity at 7 FP/scan rates (i.e. 0.125, 0.25, 0.5, 1, 2, 4, 8) was computed. The overall score of system (Competition Performance Metric-CPM) was defined as the average sensitivity for\nthese 7 FP/scan rates.",
    "paper_id": "1805.02279",
    "title": "S4ND: Single-Shot Single-Scale Lung Nodule Detection",
    "authors": [
      "Naji Khosravan",
      "Ulas Bagci"
    ],
    "published_date": "2018-05-06",
    "primary_category": "cs.CV",
    "arxiv_url": "http://arxiv.org/abs/1805.02279v2",
    "chunk_index": 15,
    "total_chunks": 20,
    "char_count": 484,
    "word_count": 69,
    "chunking_strategy": "semantic"
  },
  {
    "chunk_id": "a650f3de-aae4-40a6-91b1-3300384782bf",
    "text": "Building blocks of S4ND and comparisons: This subsection explains\nhow we build the proposed S4ND network and provides a detailed comparison\nwith several baseline approaches. We compared performance of S4ND with stateof-the-art algorithms, including SSD (single-shot multi-box object detection) [7],\nknown to be very effective for object detection in natural scenes. We show that\nSSD suffers from low performance in lung nodule detection, even though trained\nfrom scratch on LUNA dataset. A high degree of scale bias and known difficulties\nof the lung nodules detection (texture, shape, etc.) in CT data can be considered\nas potential reasons. To address this poor performance, we propose to replace\nthe convolution layers with dense blocks to improve the information flow in the\nnetwork.",
    "paper_id": "1805.02279",
    "title": "S4ND: Single-Shot Single-Scale Lung Nodule Detection",
    "authors": [
      "Naji Khosravan",
      "Ulas Bagci"
    ],
    "published_date": "2018-05-06",
    "primary_category": "cs.CV",
    "arxiv_url": "http://arxiv.org/abs/1805.02279v2",
    "chunk_index": 16,
    "total_chunks": 20,
    "char_count": 787,
    "word_count": 120,
    "chunking_strategy": "semantic"
  },
  {
    "chunk_id": "29636e50-cef3-473e-ac7e-e5b0d43f5ef5",
    "text": "Further, we experimentally tested the effects of various down sampling\ntechniques. Table 1 shows the results of different network architectures along\nwith the number of parameters based on these combinations. We implemented\nthe SSD based architecture with 3 different pooling strategies: (1) average pooling (2D Dense Avepool), (2) replacing pooling layers with convolution layers\nwith kernel size 3 × 3 and stride 2 (2D Dense Nopool) and (3) max pooling (2D\nDense Maxpool). Our experiments show that max pooling is the best choice of\nfeature sampling for tiny object detection as it selects the most discriminating\nfeature in each step. 2D Dense Nopool outperforms the normal average pooling\n(2D Dense Avepool) as it is in concept a learnable averaging over 3 × 3 regions\nof our network, based on the way we defined kernel size and stride. 3D Networks, growth rate (GR), and comparisons: We implemented\nS4ND in a completely 3D manner. Growth rate for all the blocks inside the network was initially fixed to 16 (3D Dense). However, we observed that increasing\nthe growth rate in the last 2 blocks of our network, where the computational\nexpense is lowest, (from 16 to 32 and 64, respectively) improved the performance of detection (3D Increasing GR in Table 1). Also, having deeper blocks,\neven with a fixed growth rate of 16 for all the blocks, help the information flow\nin the network and improved the results further (3D Deeper Blocks in Table\n1). The final proposed method benefits from both deeper blocks and increasing\ngrowth rate in its last two blocks. Fig. 3 (left) shows the FROC comparison of\nproposed method with the baselines. The 10-fold cross validation results were\ncompared with the current state of the art lung nodule detection method (3D\nDCNN which is the best published results on LUNA dataset) [1]. Our proposed\nmethod outperformed the best available results both in sensitivity and FROC\nscore, while only using as less as a third of its parameters, and without the need\nfor multi-stage refinements. Major findings: (1) We obtained 0.897 FROC rate in 10-fold cross validation, and consistently outperformed the state of the art methods as well as other\nalternatives. (2) SSD (the state of the art for object detection in natural images) resulted in the lowest accuracy in all experiments. Proposed S4ND, on the\nother hand, showed that single scale single shot algorithm performs better and Model Sensitivity% Num of parameters CPM\n1-fold 2D2D SSDDense Avepool 77.8%84.8% 59,790,78767,525,635 0.6490.653\n2D Dense Nopool 86.4% 70,661,955 0.658\n2D Dense Maxpool 87.5% 67,525,635 0.672 selected 3D Dense 93.7% 694,467 0.882\n3D Increasing GR 95.1% 2,429,827 0.890\n3D Deeper Blocks 94.2% 1,234,179 0.913 Randomly Proposed (S4ND) 97.2% 4,572,995 0.931\n10-fold 3DProposedDCNN(S4ND)[1] 95.2%94.6% 11,720,0324,572,995 0.8970.891",
    "paper_id": "1805.02279",
    "title": "S4ND: Single-Shot Single-Scale Lung Nodule Detection",
    "authors": [
      "Naji Khosravan",
      "Ulas Bagci"
    ],
    "published_date": "2018-05-06",
    "primary_category": "cs.CV",
    "arxiv_url": "http://arxiv.org/abs/1805.02279v2",
    "chunk_index": 17,
    "total_chunks": 20,
    "char_count": 2841,
    "word_count": 452,
    "chunking_strategy": "semantic"
  },
  {
    "chunk_id": "52154d88-97f7-4dcb-9405-a5d3537fc2d0",
    "text": "Comparison of different models with varying conditions. 0.6 0.6\n2D_SSD:0.649\n2D_Dense_Avepool:0.653 Sensitivity 0.4 2D_Dense_Nopool:0.658 Sensitivity 0.4\n2D_Dense_Maxpool:0.672\n3D_Dense:0.882\n0.2 3D_Increasing GR:0.89 0.2\n3D_Deeper blocks:0.913 State of art:0.891\nProposed:0.931 Proposed:0.897\n0.00.125 0.25 0.5 1 2 4 8 0.00.125 0.25 0.5 1 2 4 8\nAverage number of false positives per scan Average number of false positives per scan Comparison of base line as well as comparison with the state of the art. Numbers\nin front of each method in the legend show Competition Performance Metric (CPM).",
    "paper_id": "1805.02279",
    "title": "S4ND: Single-Shot Single-Scale Lung Nodule Detection",
    "authors": [
      "Naji Khosravan",
      "Ulas Bagci"
    ],
    "published_date": "2018-05-06",
    "primary_category": "cs.CV",
    "arxiv_url": "http://arxiv.org/abs/1805.02279v2",
    "chunk_index": 18,
    "total_chunks": 20,
    "char_count": 593,
    "word_count": 85,
    "chunking_strategy": "semantic"
  },
  {
    "chunk_id": "dea95610-9824-42ac-94c1-097ff8384ac8",
    "text": "more suited to tiny object detection problem. (3) The proposed method achieved\nbetter sensitivity, specificity, and CPM in single fold and 10-fold throughout experiments where S4ND used less than the half parameters of 3D DCNN (current\nstate of the art in lung nodule detection). (4) A careful organization of the\narchitecture helps avoiding computationally heavy processing. We have shown\nthat maxpooling is the best choice of feature selection throughout the network\namongst current available methods. (5) Similarly, dense and deeper connections\nimprove the detection rates through better information flow through layers. It\nshould be noted that the runtime of our algorithm for the whole scan, on the\ntest phase, varies from 11 secs to 27 secs based on the number of slices in the\nscan on a single NVIDIA TITAN Xp GPU workstation with RAM of 64 GBs. This paper introduces a single-shot single-scale fast lung nodule detection algorithm without the need for additional FP removal and user guidance for re- finement of detection process.",
    "paper_id": "1805.02279",
    "title": "S4ND: Single-Shot Single-Scale Lung Nodule Detection",
    "authors": [
      "Naji Khosravan",
      "Ulas Bagci"
    ],
    "published_date": "2018-05-06",
    "primary_category": "cs.CV",
    "arxiv_url": "http://arxiv.org/abs/1805.02279v2",
    "chunk_index": 19,
    "total_chunks": 20,
    "char_count": 1038,
    "word_count": 165,
    "chunking_strategy": "semantic"
  },
  {
    "chunk_id": "2652ce39-20b9-4ef3-813c-7108438a599c",
    "text": "Our proposed deep network structure is fully 3D\nand densely connected. We also critically analyzed the role of densely connected\nlayers as well as maxpooling, average pooling and fully convolutional down sampling in detection process. We present a fundamental solution to address the major challenges of current region proposal based lung nodule detection methods:\ncandidate detection and feature resampling stages. We experimentally validate\nthe proposed network's performance both in terms of accuracy (high sensitivity/specificity) and efficiency (less number of parameters and speed) on publicly\navailable LUNA data set, with extensive comparison with the natural object detector networks as well as the state of the art lung nodule detection methods. A\npromising future direction will be to combine diagnosis stage with the detection.",
    "paper_id": "1805.02279",
    "title": "S4ND: Single-Shot Single-Scale Lung Nodule Detection",
    "authors": [
      "Naji Khosravan",
      "Ulas Bagci"
    ],
    "published_date": "2018-05-06",
    "primary_category": "cs.CV",
    "arxiv_url": "http://arxiv.org/abs/1805.02279v2",
    "chunk_index": 20,
    "total_chunks": 20,
    "char_count": 839,
    "word_count": 122,
    "chunking_strategy": "semantic"
  }
]