Geoinformatics Team Seminar (Talk by Pengyu Dai, Institute of Science Tokyo).

Title: Towards Annotation-Efficient Deep Learning for Computer-Aided Diagnosis

Abstract:
In medical AI, the reliance on large, annotated datasets often hinders the scalability and real-world deployment of diagnostic systems. This talk introduces two complementary research directions aimed at reducing annotation dependency while preserving clinical utility. The first explores how meaningful visual representations can be learned directly from unlabeled imaging data, enabling early disease localization with minimal manual labels. The second addresses this principle to the multimodal setting, where vision-language models are enhanced through retrieval and reasoning—without requiring annotated rationales. Together, these efforts outline a unified vision for annotation-efficient AI, where both intra-modal patterns and cross-modal knowledge are leveraged to build robust, interpretable, and scalable diagnostic tools.