Classification of medical images in the context of health care is particularly challenging if there is a time-constraint for supporting diagnosis -- avoiding long delays associated with image analysis and annotation. Our goal is to provide medical images to the physicians' desktops, soon after MRI scans are acquired. The task involves predicting a cohort class for the heretofore unseen patient (and related images) and offering linkages to historical diagnosis data associated with the members of the predicted cohort class. The basic idea is to offer a relatively accurate cohort class for a new patient so that the cohort can be used as a baseline to understand current patient's status and develop a treatment plan. In this paper, we describe a system and related techniques for automatically identifying cohort classes for MRI scans. We use a set of well-curated images provided by the ADNI project for training and testing. We build upon past approaches such as reinforcement learning and relevance feedback to refine a multilevel classification system, with the aim of adapting to new information over time. The paper presents results from a large set of experiments conducted to evaluate various key system parameters, image characteristics, and relevance feedback. Overall, it is demonstrated that the proposed approach to MRI image classification has promise in terms of both effectiveness and efficiency.
展开▼
