The goal of this proposal is the automated classification of imaging studies of patients with tumors. As the role of imaging becomes increasingly important in medical care, effective methods for storing and retrieving key images will become critical. Image classification and subsequent summarization proffers a method to compress imaging studies by selecting only pertinent image slices that objectively document a patient's condition, while preserving the full integrity of the original data; as such, its applications include multimedia electronic medical records, telemedicine, and teaching files. This proposal details an innovative method to accomplish image classification based on principal component analysis. A training set of images classified by experts will be used to generate a basis set of images that captures the variance among the images. The projection on this basis set of images, called eigenimages, is used as an image index for classification and retrieval. Two key aspects critical to the success of accurate image classification are described: normalization of both image spatial and intensity properties. A modification to this methodology is also proposed to handle images with small abnormalities: image sub-regions that are 'abnormal' are located by searching the query image for the region that best matches a training set of sub-images of 'abnormal regions'. The target domain for the proposal is MR imaging studies of patients with brain tumors; in future work, this research will be extended to cover other neurological conditions, imaging modalities, and anatomical regions. Technical evaluation will be performed by comparing the automated methods with that of experts.
