MedSigLIP

A medical-grade vision-language model leveraging SigLIP architecture to align clinical imagery with precise anatomical descriptions.

MedSigLIP adapts Google’s Sigmoid Loss for Language-Image Pre-training (SigLIP) to the healthcare domain, outperforming traditional CLIP-based models in zero-shot medical classification. By training on specialized datasets like ROCO and MedICAT, the model achieves superior semantic alignment between complex visual features (radiographs or pathology slides) and technical medical nomenclature. It utilizes a ViT-B/16 backbone to process high-resolution inputs, enabling clinicians to automate image tagging and retrieval with significantly lower computational overhead than previous contrastive learning frameworks.

https://huggingface.co/flaviogdf/medsiglip-vit-patch14-224