MICCAI Industrial Talk: Optimizing Ultrasound Image Similarity Search - Jan 6. 2026 скачать в хорошем качестве

MICCAI Industrial Talk: Optimizing Ultrasound Image Similarity Search - Jan 6. 2026

Title: Optimizing Ultrasound Image Similarity Search: From Model Benchmarking to Interactive Multimodal Enhancement Speaker: Richárd Zsámboki, Senior Staff Data Scientist, GE Healthcare Abstract: This talk presents a dual perspective on advancing ultrasound image similarity search for clinical applications. First, we share a rigorous performance evaluation of state-of-the-art AI encoders (e.g., DINOv2, SAM2, DreamSim) across fetal and early pregnancy datasets, highlighting their strengths in organ and diagnosis retrieval. Then, we introduce an interactive search workflow that integrates user-provided textual annotations (view and finding labels) with image features, significantly boosting diagnostic match rates — from 38% to 75% in our tests. Together, these results demonstrate how combining robust model selection with user-guided multimodal input can transform similarity search into a clinically valuable tool for annotation, decision support, and workflow optimization. Biography: Richárd Zsámboki is a Senior Staff Data Scientist at GE Healthcare with over a decade of experience in deep learning research and development. He leads projects focused on AI applications in healthcare, including organ-at-risk segmentation, tumor detection, and ultrasound-based decision support. His recent work centers on leveraging foundation models and multimodal approaches to optimize ultrasound image similarity search for clinical workflows—the topic of his MICCAI industrial talk. Richárd believes that applying AI in medicine is one of the most impactful uses of this technology, bringing its benefits directly to patients whose lives depend on accurate and timely care.