LiangRenjie/kidney-3dct-lesion-characterization

Kidney 3D-CT Lesion Characterization — runnable data

Companion data for the code repository **kidney-3dct-lesion-characterization** (paper: Multi-Granularity 3D Kidney Lesion Characterization from CT Volumes).

The private UF Health dataset used in the paper cannot be released (patient privacy). This repository instead provides two small, fully runnable datasets so the published code can be executed end-to-end:

These datasets demonstrate that the code runs and document the on-disk data contract. They do not reproduce any number in the paper — those require the private UF Health cohort.

Layout

synthetic/
└── UF_Kidney_CT/
    ├── final_dataset/
    │   ├── train/*.npz            # keys: "image" (HU), "mask" (uint8 0-6)
    │   ├── valid/*.npz
    │   ├── study_id_to_label.json
    │   └── labels.jsonl           # L1/L2/L3 labels (see code DATA.md)
    ├── resized_train_images/*.nii.gz   # full volumes for the L1 input path
    └── resized_valid_images/*.nii.gz

kits23_subset/
├── step4_cropped/case_XXXXX.npz  # 6 cases, our cropped NPZ format
├── labels.json                   # L1/L2 labels derived from KiTS23 masks
└── ATTRIBUTION.md                # required KiTS23 citation + license

Usage

Clone the code repo, then point the environment variables at the downloaded data:

# Synthetic data, full pipeline including per-lesion L3:
KIDNEY_DATA_ROOT=synthetic/UF_Kidney_CT/final_dataset \
  python train.py --config-name smoke

# KiTS23 external validation, side-level L1/L2:
KITS23_ROOT=kits23_subset \
  python eval_kits23.py --checkpoint <path/to/best_model.pt> --label_level L3_D --mask_strategy binary

The synthetic/ set can also be regenerated from scratch (no download needed) with python tools/generate_synthetic_data.py in the code repo.

Licensing

• —synthetic/ — procedurally generated by the released code; contains no real patient data. Usable under the code repository's MIT license.
• —kits23_subset/ — derived from the KiTS23 dataset, which is released under CC BY-NC-SA 4.0 by the KiTS challenge organizers. The ShareAlike and NonCommercial terms apply to this subset and any derivative of it. You must cite KiTS23 — see kits23_subset/ATTRIBUTION.md. The full dataset is at <https://kits-challenge.org/kits23/>.