wq2012/knee_3d_mri_segmentation_SKI10

Knee Bone & Cartilage Segmentation Models (Baseline v1 SKI10)

This repository contains the baseline v1 trained models for 3D MRI Knee Bone and Cartilage Segmentation on the SKI10 dataset, using the kneeseg Python package.

Recommended Newer Models: For substantially higher accuracy (0.9689 Femur, 0.9742 Tibia, 0.8026 Femoral Cartilage, 0.8001 Tibial Cartilage DSC on the strict 80/20 non-overlapping split) and 3.6x smaller checkpoints (1.56 GB vs. 5.67 GB), see wq2012/knee_3d_mri_segmentation_SKI10_subchondral_atlas or the unified cross-domain model wq2012/knee_3d_mri_segmentation_SKI10_OAI_subchondral_atlas.

• GitHub: https://github.com/wq2012/kneeseg
• PyPI: https://pypi.org/project/kneeseg/
• Hugging Face Collection: Knee MRI Segmentation Collection
• Sibling Models:
  • wq2012/knee_3d_mri_segmentation_SKI10_subchondral_atlas (Recommended SKI10 4-structure model)
  • wq2012/knee_3d_mri_segmentation_OAI_downsampled_subchondral_atlas (Recommended OAI 6-structure model)
  • wq2012/knee_3d_mri_segmentation_SKI10_OAI_subchondral_atlas (Unified Cross-Domain SKI10 + OAI model)
  • wq2012/knee_3d_mri_segmentation_SKI10 (This repository — Baseline v1 SKI10 model)
  • wq2012/knee_3d_mri_segmentation_OAI_downsampled (Baseline v1 OAI downsampled model)

Model Details

• Architecture: Dense Random Forest with Auto-Context (2 Passes for Bones, 2 Passes for Cartilage).
• Dataset:
  • Training: The models in this repository were trained on all 100 cases from the SKI10/TrainingData folder.
  • Evaluation: Evaluated on a subset of 20 cases of SKI10/TrainingData (note: in this v1 full-100 setup, those 20 cases are included in the 100 training scans; on the strict 80-train / 20-eval non-overlapping split in ski10_full_split.json, the baseline v1 architecture achieves 0.9046 ± 0.0361 Femur, 0.9292 ± 0.0260 Tibia, 0.6767 ± 0.0481 Femoral Cartilage, and 0.6411 ± 0.0540 Tibial Cartilage DSC).
  • Inference: Predicted labels for the 50 official SKI10 TestData cases are included in the TestData_labels/ directory.
• Performance:
  • Strict 80/20 Non-Overlapping Split (20 Held-Out Cases):
    • Femur Bone: 0.9046 ± 0.0361 DSC
    • Tibia Bone: 0.9292 ± 0.0260 DSC
    • Femoral Cartilage: 0.6767 ± 0.0481 DSC
    • Tibial Cartilage: 0.6411 ± 0.0540 DSC
  • 100-Train In-Sample Subset (20 Cases with Train-Eval Overlap):
    • Femur Bone: 0.9155 ± 0.0303 DSC
    • Tibia Bone: 0.9383 ± 0.0194 DSC
    • Femoral Cartilage: 0.7095 ± 0.0478 DSC
    • Tibial Cartilage: 0.6799 ± 0.0429 DSC

Usage

These models are designed to be used with the kneeseg library.

1. Install Library

pip install --upgrade kneeseg

2. Load Models

from kneeseg.bone_rf import BoneClassifier
from kneeseg.rf_seg import CartilageClassifier

# Load Bone Model (Pass 1)
bone_p1 = BoneClassifier()
bone_p1.load("bone_rf_p1.joblib")

# ... (See library documentation for full pipeline)

Files

• bone_rf_p1.joblib (3.70 GB): Bone Segmentation Pass 1 (Voxel Classifier)
• bone_rf_p2.joblib (734.22 MB): Bone Segmentation Pass 2 (Auto-Context Refinement)
• cartilage_rf_p1.joblib (923.73 MB): Cartilage Segmentation Pass 1 (Semantic Context Forest)
• cartilage_rf_p2.joblib (305.17 MB): Cartilage Segmentation Pass 2 (Auto-Context Refinement)
• TestData_labels/: Predicted labels of the SKI10 official TestData using our models.

Citation

Plain Text:

Quan Wang, Dijia Wu, Le Lu, Meizhu Liu, Kim L. Boyer, and Shaohua Kevin Zhou. "Semantic Context Forests for Learning-Based Knee Cartilage Segmentation in 3D MR Images." MICCAI 2013: Workshop on Medical Computer Vision.

Quan Wang. Exploiting Geometric and Spatial Constraints for Vision and Lighting Applications. Ph.D. dissertation, Rensselaer Polytechnic Institute, 2014.

BibTeX:

@inproceedings{wang2013semantic,
  title={Semantic context forests for learning-based knee cartilage segmentation in 3D MR images},
  author={Wang, Quan and Wu, Dijia and Lu, Le and Liu, Meizhu and Boyer, Kim L and Zhou, Shaohua Kevin},
  booktitle={International MICCAI Workshop on Medical Computer Vision},
  pages={105--115},
  year={2013},
  organization={Springer}
}

@phdthesis{wang2014exploiting,
  title={Exploiting Geometric and Spatial Constraints for Vision and Lighting Applications},
  author={Quan Wang},
  year={2014},
  school={Rensselaer Polytechnic Institute}
}