• Bonnet: Ultra-Fast Whole-Body Bone Segmentation from CT Scans

  Bonnet is an ultra-fast whole-body bone segmentation pipeline for CT scans. It runs in seconds per scan on a single commodity GPU while maintaining reliable segmentation quality across different datasets.

  ---

  Contents

  • Links: Processed training data & model weights
  • Data Preparation
  • Training
  • Inference (Single Case vs. Sample Data)

  ---

  Links: Processed training data & model weights

  • Processed training data: Processed Data
  • Model weights: Weights

  ---

  Data Preparation

  If you want to train on your own dataset, use the provided preprocessing script to convert CT and segmentation files into sparse voxel formats (.npz).

  1. Organize your raw data

  Ensure your data follows this structure:

  DS_ORIGINAL/
  ├── s0001/
  │   ├── ct.nii.gz
  │   └── segmentations/ 
  │       ├── vertebra_C1.nii.gz
  │       └── ...
  └── s0002/
  

  2. Run Preprocessing

  The script src/preprocess_dataset.py extracts voxels within the [200, 3000] HU range and maps labels according to conf/classes.yaml.

  Bash

  python src/preprocess_dataset.py
  

  • Input: Defined by INPUT_ROOT in the script.
  • Output: Generates sparse_voxel.npz for each case in the OUTPUT_ROOT.

  ---

  Training

  1. Configure Paths:
     • Open Bonnet/conf/data/totalseg_hu200_3000.yaml.
     • Set dataset_path and cache_path to your preprocessed data folder.
  2. Run Training:

  python main.py
  

  ---

  Inference

  Option A: Single Case Inference (End-to-End)

  Use inference_single.py to segment a raw .nii.gz CT scan directly. This script handles preprocessing, windowed inference, and volume reconstruction automatically.

  Bash

  python inference_single.py --ct /path/to/ct.nii.gz --out ./prediction.nii.gz --device cuda
  

  Note: Ensure the CHECKPOINT_PATH and CONFIG_PATH inside the script point to your actual local files.

  Option B: Evaluation with Sample Data

  To run evaluation on the provided sample test set:

  1. Open Bonnet/conf/eval/eval_on_test.yaml and set eval_only: True.
  2. Point your data config to the sample data path.
  3. Run: python main.py