Building a CBIR Benchmark with TotalSegmentator and FAISS

This article explores the methods and datasets used to build a benchmark for content-based image retrieval (CBIR) in medical imaging. It examines vector databases, the challenges of large-scale similarity search, and indexing techniques such as flat search, Locality Sensitive Hashing (LSH), and Hierarchical Navigable Small World (HNSW). The Facebook AI Similarity Search (FAISS) library is used to implement efficient approximate nearest neighbor (ANN) search. Using the TotalSegmentator dataset of over 1,200 CT volumes, embeddings were extracted slice-by-slice and indexed, enabling rapid, metadata-free retrieval across more than 290,000 image embeddings.