Benchmarking Foundation Models for Renal Lesion Stratification in CT

Hartmut Häntze, Sarah De Boer, Myrthe Buser, Alessa Hering, Bram van Ginneken, Mathias Prokop, Jawed Nawabi, Sebastian Ziegelmayer, Lisa C. Adams, Keno Bressem

Keywords: Computed Tomography, Foundation Model, Renal Cell Carcinoma

TL;DR: We investigated whether medical foundation models can advance classification performance for the data-scarce task of CT-based renal lesion stratification.

Registration Requirement: Yes

Abstract: The rapid proliferation of open-source medical foundation models (FMs) raises a practical question: how well do their pre-trained representations transfer to clinically relevant but data-scarce classification tasks? We investigated this in a controlled benchmark of three medical FMs on renal lesion stratification, a multi-class problem, where training data is inherently scarce. Using a frozen feature-probing protocol, we compared FM embeddings against two established baselines: a radiomics classifier and a 3D ResNet-50 trained from scratch. Our results reveal two findings. First, FM performance matched the from scratch-trained ResNet, while drastically lowering the hardware demand. Second, the conventional radiomics baseline (AUC 0.88) outperformed all deep learning approaches (AUC [0.69, 0.77]), suggesting that, despite their potential to improve classification in data-scarce settings, medical FMs do not yet surpass established models for renal lesion stratification.

Reproducibility: https://anonymous.4open.science/r/RenalVision-8916

Visa & Travel: No

Read CFP & Author Instructions: Yes

Originality Policy: Yes

Single-blind & Not Under Review Elsewhere: Yes

LLM Policy: Yes

Submission Number: 69