MuCTaL

Brian Isett, Rebekah Dadey, Aofei Li, Ryan C. Augustin, Kate Smith, Aatur D. Singhi, Qiangqiang Gu, Riyue Bao. A lightweight multi-cancer tumor localization framework for whole-slide histopathology images (H&E). DOI:10.48550/arXiv.2603.08844.

MuCTaL provides scripts and notebooks for end-to-end pathology AI workflows:

• whole-slide preprocessing and tiling,
• tile-level model training and inference (FastAI/PyTorch),
• post-processing into clinically useful outputs (heatmaps and GeoJSON regions),
• cross-validation utilities and exploratory notebooks.

Trained model available on huggingface https://huggingface.co/hillmancancercenterds/MuCTaL

---

Repository Summary

This repository is organized around a practical WSI (whole slide image) pipeline:

1. Preprocess slides into tissue-rich tiles using pathml-based pipelines.
2. Train CNN models on labeled tiles (FastAI).
3. Run inference on unseen tiles/slides.
4. Convert predictions to visual artifacts (heatmaps) and annotation formats (GeoJSON).
5. Evaluate folds/repeats with cross-validation helper scripts.

The codebase appears to be research/HPC-oriented and includes path placeholders and batch-style scripts, so users should expect to adapt data paths and execution wrappers to their environment.

---

Folder Structure

MuCTaL/
├── LICENSE
├── README.md
├── helpers/
│   ├── __init__.py
│   ├── anno.py
│   ├── preproc.py
│   └── tile.py
├── notebooks/
│   ├── 01_generate_wsi_samplesheet_run_preprocessing.ipynb
│   ├── 02_annotated_tile_file_org_for_training.ipynb
│   ├── 03_train_model_fastai2.7.ipynb
│   ├── 04_model_eval_fastai2.7.ipynb
│   ├── 05_example_inference_to_geojson.ipynb
│   ├── 06_acral_tile_heatmap_class_viz.ipynb
│   └── 07_percent_predicted_tumor_each_slide.ipynb
├── pipeline/
│   ├── fastai_inference_v10.py
│   ├── pathml_preproc_v10.py
│   ├── tile_infer_to_geojson.py
│   └── tile_infer_to_heatmap.py
└── train/
    └── train_full.py

---

What Each Main Module Does

• pipeline/: Main runnable pipeline scripts for preprocessing, inference, and output generation.
• train/: End-to-end model training script for full dataset training.
• helpers/: Reusable utility code (annotation geometry checks, preprocessing helpers, tile parsing).
• notebooks/: Interactive analysis/tutorial notebooks for data prep, training, evaluation, and visualization.

---

Typical Workflow

1. Prepare sample metadata (notebooks and TSV inputs).
2. Preprocess WSIs to tiles with pipeline/pathml_preproc_v10.py.
3. Build balanced tile CSVs
4. Train model(s) with scripts in train/
5. Infer tile probabilities using pipeline/fastai_inference_v10.py.
6. Generate outputs:
   • GeoJSON tumor regions: pipeline/tile_infer_to_geojson.py
   • Slide heatmaps/overlays: pipeline/tile_infer_to_heatmap.py

---

Requirements

Core dependencies inferred from scripts:

• Python 3.9+
• fastai, torch, torchvision
• pathml
• opencv-python
• numpy, pandas, scipy, matplotlib, tqdm
• dask, distributed (optional, for distributed preprocessing)
• Pillow
• cv2geojson

Many scripts are designed for HPC/SLURM environments and may rely on environment variables (e.g., SLURM_SCRATCH) and local data layouts.

---

Quick Start

Update all paths below for your environment.

1) Preprocess a slide to tiles

python pipeline/pathml_preproc_v10.py \
  /path/to/output \
  /path/to/slide.svs \
  224 \
  /path/to/MuCTaL

2) Run tile inference

python pipeline/fastai_inference_v10.py \
  /path/to/tiles_df.tsv \
  /path/to/model.pkl \
  /path/to/tile_root/ \
  /path/to/output

3) Convert predictions to GeoJSON

python pipeline/tile_infer_to_geojson.py \
  /path/to/infer_tiles.tsv \
  /path/to/output_geojson

4) Generate heatmap overlay

python pipeline/tile_infer_to_heatmap.py \
  /path/to/infer_tiles.tsv \
  /path/to/original_slide.svs \
  /path/to/output_heatmaps

---

Notes and Caveats

• Several scripts contain hard-coded placeholders (e.g., /path/to/...) and versioned naming conventions.
• Some utilities expect specific input TSV schemas (tile paths, slide/case IDs, class labels).
• Cross-validation scripts assume specific model naming patterns such as: arch_kfold_nrep_nbal_px_v (example style from code).

---

License

This project includes a LICENSE file in the repository root. See that file for usage terms.