Foundation model embeddings for quantitative tumor imaging biomarkers.

Area of Science:

• Artificial Intelligence in Medical Imaging
• Quantitative Radiomics
• Foundation Models

Background:

• Foundation models are increasingly adopted in medical imaging.
• Systematic evaluation of their ability to extract reliable quantitative radiographic phenotypes across diverse clinical contexts is lacking.

Purpose of the Study:

• Introduce TumorImagingBench, a benchmark for evaluating medical imaging foundation models.
• Assess model performance in deriving deep learning-based radiographic phenotypes for cancer.
• Compare model robustness, interpretability, and embedding similarity.

Main Methods:

• Curated a benchmark (TumorImagingBench) with six public datasets (3,244 scans) and varied oncological endpoints.
• Evaluated ten medical imaging foundation models (2020-2025) on performance, robustness, and interpretability.
• Compared mutual similarity of learned embedding representations.

Main Results:

• Revealed performance disparities among evaluated foundation models.
• Demonstrated differences in robustness to common sources of variability.
• Showcased variations in saliency-based interpretability and embedding similarity.

Conclusions:

• The benchmark provides critical insights for selecting optimal foundation models for quantitative imaging tasks.
• Highlights the need for systematic evaluation of AI models in medical imaging.
• Public release of code, datasets, and results promotes reproducible research.