FailCatcher 2.0.0

Post-hoc uncertainty quantification toolkit for PyTorch deep learning models

FailCatcher

FailCatcher is a uncertainty quantification (UQ) toolkit for PyTorch classification deep learning models, developed and benchmarked on medical imaging datasets from the MedMNIST collection and external test sets. Failure detection benchmark results can be found in the benchmarks README.

Preprint: Steinmetz et al., medRxiv 2026 — DOI: 10.64898/2026.05.04.26350496

The project provides:

• A reusable Python library (ToolBox/) implementing multiple UQ methods with a clean, unified API.
• A full benchmarking pipeline (Benchmarks/), including model training, classification evaluation and failure detection evaluation on a diverse set of distribution shifts images.

---

Repository structure

FailCatcher/
├── ToolBox/                        # UQ library (installable Python package)
│   ├── failure_detection.py        # High-level FailureDetector API
│   ├── UQ_toolbox.py               # Public API aggregator
│   ├── core/                       # Base classes and shared inference utilities
│   ├── methods/                    # UQ method implementations
│   │   ├── tta.py                  # Test-Time Augmentation (TTA) and GPS
│   │   ├── ensemble.py             # Ensemble STD and MC Dropout
│   │   ├── distance.py             # MSR, MLS, and calibration methods
│   │   └── latent.py               # KNN and SHAP latent-space methods
│   ├── search/                     # Greedy Policy Search (GPS) algorithm
│   ├── evaluation/                 # AUROC, AURC, AUGRC metrics and plots
│   ├── visualization/              # Visualization utilities
│   └── tests/                      # Smoke tests and pre-run checks
│
├── Benchmarks/
│   └── medMNIST/
│       ├── launcher_benchmark.py   # Top-level benchmark launcher
│       ├── run_medmnist_benchmark.py # Core benchmark runner (single config)
│       ├── trainings/              # Model training scripts and launchers
│       ├── utils/                  # Data loading, preprocessing, visualization
│       │   ├── train_models_load_datasets.py  # Central data/model utilities
│       │   ├── dataset_utils.py               # External datasets, corruptions
│       │   └── data_preprocessing_classification_evaluation/
│       ├── data/                   # External test datasets (AMOS-2022, MIDOG++, ISIC)
│       ├── models/                 # Trained model checkpoints
│       ├── runs/                   # Training logs and per-run artifacts
│       └── results/                # Benchmark outputs (JSON, figures, cache)
│
└── requirements.txt

---

Quick start

1. Install dependencies

pip install -r requirements.txt

For covariate-shift corruption benchmarks, also install:

pip install medmnistc

2. Install the ToolBox library

cd ToolBox
pip install -e .

3. Download models and datasets from HuggingFace

Pre-trained model checkpoints and pre-processed external datasets are available on HuggingFace. Run the one-command setup to skip training and manual preprocessing:

python scripts/setup_from_hub.py

• Models (~59 GB, 325 checkpoints): pstnmz/FailCatcher-models
• Datasets (AMOS-2022, MIDOG++, DermaMNIST-E): pstnmz/FailCatcher-datasets

You can download models or datasets independently:

python scripts/setup_from_hub.py --models-only
python scripts/setup_from_hub.py --datasets-only --datasets amos22 midog dermamnist-e

4. Train models (alternative to step 3)

See Benchmarks/README.md for the full reproducible training and benchmarking pipeline.

5. Run the benchmark

python Benchmarks/medMNIST/launcher_benchmark.py \
    --python /path/to/your/venv/bin/python \
    --datasets breastmnist organamnist \
    --models resnet18 \
    --setups "" DA \
    --gpu 0

---

UQ methods

Method	Description
MSR	Maximum Softmax Response — distance between predicted probability and 1
MSR-calibrated	MSR after temperature / Platt scaling calibration
MLS	Maximum Logit Score — pre-softmax equivalent of MSR
Ensembling	Standard deviation across 5-fold CV model predictions
TTA	Test-Time Augmentation — std over random augmentation passes
GPS	Greedy Policy Search — optimised TTA policy found on the calibration set
KNN-Raw	k-NN distance in avgpool latent space
KNN-SHAP	KNN with SHAP-weighted latent features
MC Dropout	Monte Carlo Dropout at inference time
ZScore Aggregation	Z-score normalised aggregation of multiple methods

---

Datasets

Internal test sets (MedMNIST)

breastmnist, pneumoniamnist, organamnist, octmnist, pathmnist, bloodmnist, tissuemnist, dermamnist-e

External test sets (not in git — see Benchmarks/README.md for setup)

• AMOS-2022 — abdominal CT organ patches mapped to OrganaMNIST classes. Available on HuggingFace or preprocessed via data/AMOS_2022/read_npz.ipynb.
• MIDOG++ — mitosis detection histology patches as OOD test for PathMNIST. Available on HuggingFace or generated by utils/data_preprocessing_classification_evaluation/create_midog_patch_dataset.py.
• DermaMNIST-E — extended DermaMNIST with ID and external centre splits. Available on HuggingFace or downloaded from Zenodo, loaded by utils/data_preprocessing_classification_evaluation/local_dermamnist_e.py.

---

Reproducibility

All benchmark results are reproducible from scratch:

• Random seeds fixed to 42 everywhere (training, CV splits, TTA).
• 5-fold StratifiedKFold CV with seed=42 is consistent between training and inference.
• Model checkpoints, result JSONs, and caches are saved with configuration-specific suffixes.
• See Benchmarks/README.md for step-by-step instructions.

---

Python version and environment

Tested with Python 3.12 and the following key packages:

Package	Version
torch	2.6.0
torchvision	0.21.0
numpy	2.1.3
scikit-learn	1.6.1
monai	1.5.1
medmnist	3.0.1
shap	0.46.0
matplotlib	3.10.0
seaborn	0.13.2

License

This project is licensed under CC BY-NC-SA 4.0.

The code is intended for research and academic use only. Commercial use is prohibited.

For commercial use, please contact the author.