Update README.md
Browse files
README.md
CHANGED
|
@@ -16,8 +16,6 @@ size_categories:
|
|
| 16 |
- 1K<n<10K
|
| 17 |
---
|
| 18 |
|
| 19 |
-
<div align="center">
|
| 20 |
-
|
| 21 |
# PPCNet Dataset
|
| 22 |
## Biplanar DRRs + Projection Matrices + Ground-Truth Point Clouds
|
| 23 |
|
|
@@ -28,11 +26,9 @@ size_categories:
|
|
| 28 |
|
| 29 |
*A curated lumbar spine dataset for 3D point cloud reconstruction from biplanar radiographs.*
|
| 30 |
|
| 31 |
-
</div>
|
| 32 |
-
|
| 33 |
---
|
| 34 |
|
| 35 |
-
##
|
| 36 |
|
| 37 |
This dataset provides paired biplanar DRRs, calibrated 3×4 projection matrices, and dense ground-truth point clouds for **1,037 patients** with complete L1–L5 lumbar vertebrae, derived from the publicly available **VerSe'19**, **VerSe'20**, and **CTSpine1K** collections.
|
| 38 |
|
|
@@ -55,7 +51,7 @@ Each of the 1,037 patient folders contains:
|
|
| 55 |
|
| 56 |
---
|
| 57 |
|
| 58 |
-
##
|
| 59 |
|
| 60 |
Fixed split in `dataset_split.json` (seed=42):
|
| 61 |
|
|
@@ -68,7 +64,7 @@ Fixed split in `dataset_split.json` (seed=42):
|
|
| 68 |
|
| 69 |
---
|
| 70 |
|
| 71 |
-
##
|
| 72 |
|
| 73 |
DRRs are generated using **Plastimatch's ray-casting algorithm** with the following parameters:
|
| 74 |
|
|
@@ -86,15 +82,15 @@ Each DRR comes with a calibrated **3×4 perspective projection matrix** encoding
|
|
| 86 |
|
| 87 |
---
|
| 88 |
|
| 89 |
-
##
|
| 90 |
|
| 91 |
The ground-truth point clouds are constructed through a five-step pipeline:
|
| 92 |
|
| 93 |
-
1.
|
| 94 |
-
2.
|
| 95 |
-
3.
|
| 96 |
-
4.
|
| 97 |
-
5.
|
| 98 |
|
| 99 |
---
|
| 100 |
|
|
@@ -120,7 +116,7 @@ huggingface-cli download ppcnet-dataset/PPCNet --repo-type dataset --local-dir .
|
|
| 120 |
|
| 121 |
---
|
| 122 |
|
| 123 |
-
##
|
| 124 |
|
| 125 |
This dataset builds upon:
|
| 126 |
|
|
@@ -129,7 +125,7 @@ This dataset builds upon:
|
|
| 129 |
|
| 130 |
---
|
| 131 |
|
| 132 |
-
##
|
| 133 |
|
| 134 |
This dataset is released under the [MIT License](LICENSE).
|
| 135 |
|
|
|
|
| 16 |
- 1K<n<10K
|
| 17 |
---
|
| 18 |
|
|
|
|
|
|
|
| 19 |
# PPCNet Dataset
|
| 20 |
## Biplanar DRRs + Projection Matrices + Ground-Truth Point Clouds
|
| 21 |
|
|
|
|
| 26 |
|
| 27 |
*A curated lumbar spine dataset for 3D point cloud reconstruction from biplanar radiographs.*
|
| 28 |
|
|
|
|
|
|
|
| 29 |
---
|
| 30 |
|
| 31 |
+
## Overview
|
| 32 |
|
| 33 |
This dataset provides paired biplanar DRRs, calibrated 3×4 projection matrices, and dense ground-truth point clouds for **1,037 patients** with complete L1–L5 lumbar vertebrae, derived from the publicly available **VerSe'19**, **VerSe'20**, and **CTSpine1K** collections.
|
| 34 |
|
|
|
|
| 51 |
|
| 52 |
---
|
| 53 |
|
| 54 |
+
## Data Split
|
| 55 |
|
| 56 |
Fixed split in `dataset_split.json` (seed=42):
|
| 57 |
|
|
|
|
| 64 |
|
| 65 |
---
|
| 66 |
|
| 67 |
+
## DRR Generation
|
| 68 |
|
| 69 |
DRRs are generated using **Plastimatch's ray-casting algorithm** with the following parameters:
|
| 70 |
|
|
|
|
| 82 |
|
| 83 |
---
|
| 84 |
|
| 85 |
+
## Ground-Truth Generation
|
| 86 |
|
| 87 |
The ground-truth point clouds are constructed through a five-step pipeline:
|
| 88 |
|
| 89 |
+
1. **Surface extraction** — marching cubes on per-vertebra segmentation masks
|
| 90 |
+
2. **Coordinate transform** — vertices mapped to physical (mm) space via NIfTI affine
|
| 91 |
+
3. **Concatenation** — all five vertebrae (L1–L5) merged into a single cloud
|
| 92 |
+
4. **Alignment** — rigid ICP with four flip initialisations to match DRR projection space
|
| 93 |
+
5. **Sampling** — farthest-point sampling to a uniform **5,120-point** representation
|
| 94 |
|
| 95 |
---
|
| 96 |
|
|
|
|
| 116 |
|
| 117 |
---
|
| 118 |
|
| 119 |
+
## Source Datasets
|
| 120 |
|
| 121 |
This dataset builds upon:
|
| 122 |
|
|
|
|
| 125 |
|
| 126 |
---
|
| 127 |
|
| 128 |
+
## License
|
| 129 |
|
| 130 |
This dataset is released under the [MIT License](LICENSE).
|
| 131 |
|