Multimodal Medical Case-Based Retrieval on the Radiology Image and Report: SNUMedinfo at VISCERAL Retrieval Benchmark

摘要

This paper describes the participation at the VISCERAL Retrieval benchmark. The task is about retrieving relevant medical cases from radiology image and report. Both query and retrieval datasets are composed of multimodal data. We extracted low-level visual features (SURF) from images and trained a query-specific SVM classifier for image retrieval. For textual retrieval, we estimated relevance with an anatomy-pathology paired RadLexID similarity function. In mixed retrieval, we combined them using weighted Borda-fuse method.