Bridged Semantic Alignment for Zero-shot 3D Medical Image Diagnosis

The Bridged Semantic Alignment (BrgSA) framework represents a significant advancement in the field of medical imaging, particularly for 3D images like computed tomography. Traditional supervised learning methods have struggled due to their reliance on extensive manual annotations, which are often limited in availability and diversity. In contrast, BrgSA leverages vision-language alignment (VLA) to facilitate zero-shot learning, allowing for improved diagnostic capabilities without the need for additional annotations. By employing a large language model for semantic summarization and a Cross-Modal Knowledge Interaction module, BrgSA effectively bridges the gap between visual and textual embeddings, enhancing their alignment. This innovative approach has been empirically validated, achieving state-of-the-art performance on a newly constructed benchmark dataset that includes 15 underrepresented abnormalities, thus paving the way for more robust and accessible medical diagnostics.