[2506.18434] Benchmarking Foundation Models and Parameter-Efficient Fine-Tuning for Prognosis Prediction in Medical Imaging

View PDF of the article “Standard Basis Models and Efficient Parameter Fine-Tuning for Diagnosis Prediction in Medical Imaging,” by Filippo Ruffini and 4 other authors

View PDF HTML (beta)

a summary:Despite the great potential of basis models (FMs) in medical imaging, their application to prognosis prediction remains challenging due to data scarcity, class imbalance, and task complexity, which limits their clinical adoption. This study provides the first systematic benchmark to evaluate the power and efficiency of transfer learning strategies for FMs compared to convolutional neural networks (CNNs) in predicting outcomes of COVID-19 patients from chest X-rays. The goal is to systematically compare fine-tuning strategies, both classical and instrumental parameters, under realistic clinical constraints related to data scarcity and class imbalance, providing empirical guidance for the deployment of AI in clinical workflow. Four publicly available COVID-19 chest X-ray datasets were used, covering mortality, severity and ICU admission, with varying sample sizes and stratification imbalance. CNNs pre-trained on ImageNet and FMs pre-trained on public or biomedical datasets were adapted using full fine-tuning, linear checking, and parameter efficient methods. Models were evaluated under full data and a few imaging systems using Matthews correlation coefficient (MCC) and recall precision AUC (PR-AUC), with cross-validation and weighted class losses. CNNs with full fine-tuning performed strongly on small, imbalanced datasets, while FM networks with efficient parameter fine-tuning (PEFT), especially LoRA and BitFit, achieved competitive results on larger datasets. Severe class imbalance degraded PEFT performance, while balanced data mitigated this effect. In low-shot settings, FMs showed limited generalization, with linear screening yielding the most stable results. No single fine-tuning strategy has proven to be universally optimal: CNNs remain reliable in low-resource scenarios, while FMs benefit from parametrically efficient methods when data is sufficient.