Medical Image Representation Learning

The field studies how models can encode useful structure from images such as CT, MRI, X-rays, ultrasound, retinal scans, and pathology slides, especially when expert labels are scarce. Common approaches include contrastive learning, masked image modeling and joint-embedding prediction, as well as vision-language alignment and anatomical or temporal supervision. Researchers test whether these representations transfer through linear probes, few-shot learning, segmentation, retrieval, and evaluation across hospitals or acquisition protocols. A central tension is that specialized pretraining can help on closely matched tasks, but performance is uneven across modalities and domains, and global features may miss small lesions or clinically important spatial detail.