Dementia diagnosis using similar and dissimilar retrieval items

Image-based disease diagnosis often requires radiologists´ qualitative interpretations that are highly dependent on their levels of expertise, and physical and mental status. Automated image analysis tools can help radiologists increase the diagnosis accuracy by providing quantitative measures. Accordingly, this paper presents an automated method for dementia diagnosis using search and retrieval of brain MR (magnetic resonance) images. The main contributions of the method are 1) its generic decision-making based on similar and dissimilar cases retrieved from a database that can be used to diagnose various brain disorders, 2) it realizes dementia diagnosis utilizing a tailored version of histogram of oriented gradients (HOGs) as features, and 3) it achieves high performance in dementia diagnosis that is independent of the database size. Comprehensive experiments with real data showed that combining information from similar and dissimilar cases leads to improved diagnostic accuracy than using similar or dissimilar cases alone, accuracy diminishes with extreme quantization of the HOGs and with small databases, and our method achieves high performance with comparable sensitivity score to the most skilled experts at better specificity figures.