Content-based image retrieval for Alzheimer's disease detection

Abstract

This paper describes ViewFinder Medicine (vfM) as an application of content-based image retrieval to the domain of Alzheimer's disease and medical imaging in general. The system follows a multi-tier architecture which provides the flexibility in experimenting with different representation, classification, ranking and feedback techniques. Classification is central to the system because besides providing an estimate of what stage of the disease the input query may belong to, it also helps adapt and rank the search results. It was found that using our multi-level approach, the classification performance matched the best result reported in the medical imaging literature. Up to 87% of patients were correctly classified in their respective classes, leading to an average precision of about 0.8 without any relevance feedback from the user. To encourage engagement and leverage physicians' knowledge, a relevance feedback function was subsequently added and as result precision improved to 0.89.