| Title |
Adapting Multimodal Foundation Models for Pathology and Radiology: Few-shot Fine-tuning Across 11 Clinical Imaging Tasks |
| DOI |
https://doi.org/10.5573/ieie.2025.62.12.93 |
| Keywords |
Foundation models; Few-shot fine-tuning; Pathology and radiology; Medical imaging; Multi-task adaptation |
| Abstract |
The emergence of multimodal foundation models pre-trained on large-scale data has opened new opportunities for medical imaging, where data scarcity and annotation costs often limit the development of task-specific solutions. In this study, we systematically investigate the adaptation of foundation models to diverse clinical applications in pathology and radiology using few-shot fine-tuning strategies. Building upon the UNICORN benchmark, we focus exclusively on 11 vision-based tasks spanning classification, detection, and segmentation across digital pathology whole-slide images and radiological modalities such as CT and MRI Our framework employs a unified backbone model with lightweight task-specific adapters, trained with limited labeled examples (“shots”) provided per task. This design allows consistent evaluation of model generalization, robustness, and efficiency across heterogeneous medical imaging scenarios. Tasks include prostate biopsy grading, lung nodule malignancy classification, cancer lesion detection in MRI, and lesion or anatomical structure segmentation in CT and MRI, among others. Experimental results demonstrate that foundation models, when paired with carefully designed few-shot adapters, achieve superior performance compared to existing baselines, while requiring substantially fewer annotated samples. Furthermore, our analysis highlights the challenges of domain shift between pathology and radiology, and the importance of adapter design in balancing accuracy with computational constraints. |