TumorImagingBench is a framework for extracting foundation model embeddings from medical images and benchmarking them across radiomics datasets.

Overview

• Unified interface for multiple foundation model extractors.
• Dataset-specific feature extraction pipelines.
• Analysis workflows in notebooks for performance, robustness, and stability.

Repository Structure

TumorImagingBench/ ├── src/tumorimagingbench/ # Core package (models, evaluation) ├── scripts/ # Utility scripts ├── tutorials/ # Tutorials and guides ├── notebooks/ # Analysis notebooks ├── data/ # Datasets (ignored by git) ├── dist/ # Large weights (ignored by git) ├── metrics/ # Evaluation outputs └── plots/ # Figures and plots 

Installation

uv sync uv run python -m pip install -e .

Python requirement: >=3.10,<3.12.

Quickstart List available extractors:

from tumorimagingbench.models import get_available_extractors print(get_available_extractors())

Load a model and initialize weights:

from tumorimagingbench.models import get_extractor Model = get_extractor("VISTA3DExtractor") model = Model() model.load()

Feature Extraction Example using the LUNA16 extractor:

uv run python src/tumorimagingbench/evaluation/luna_feature_extractor.py \ --output features/luna.pkl \ --train-csv /path/to/train.csv \ --val-csv /path/to/val.csv \ --test-csv /path/to/test.csv

Notes:

• Dataset CSVs should include image_path, coordX, coordY, coordZ (and optional labels).
• Many extractors ship with absolute default paths; override them via flags.
• Feature extraction expects a CUDA-capable GPU.

Supported Models

• CTClipVitExtractor
• CTFMExtractor
• FMCIBExtractor
• MedImageInsightExtractor
• MerlinExtractor
• ModelsGenExtractor
• PASTAExtractor
• SUPREMExtractor
• VISTA3DExtractor
• VocoExtractor

Supported Datasets

• LUNA16
• DLCS (Duke Lung Cancer Screening)
• NSCLC Radiomics
• NSCLC Radiogenomics
• C4C-KiTS
• Colorectal Liver Metastases
• LNDb
• RIDER (test-retest stability)

Tutorials

• See tutorials/README.md for guided notebooks and dataset/model integration walkthroughs.

Evaluation

• Example modelling workflow: notebooks/modelling/luna_modelling.ipynb (LUNA16 evaluation notebook).
• Loads extracted features from data/features/luna.pkl and evaluates per-model performance.
• Baselines: k-NN probing with AUC and 95% CI; linear probing (logistic regression); few-shot (1/5/10-shot).
• Visual outputs saved to plots/ (e.g., luna_auc.png, luna_knn_overlap.png, luna_few_shot.png, luna_evaluation_protocols.png).
• Ensemble methods: alignment-weighted k-NN and stacking meta-learner; weight and comparison plots (e.g., luna_alignment_weights.png, luna_stacking_weights.png, luna_ensemble_comparison.png, luna_ensemble_vs_individual.png).
• Aggregates results into overall_results.csv.

Contributing

• Follow the existing code style and update docs with changes.
• Add targeted tests for new functionality.

Citation

@article{TumorImagingBench, title={Foundation model embeddings for quantitative tumor imaging biomarkers}, author={}, journal={}, year={}, volume={}, pages={}, publisher={} }

License MIT. See LICENSE.