Similarity Classification and Retrieval in Cancer Images and Informatics

Abstract

Techniques in image similarity, classification, and retrieval of breast cancer images and informatics are presented in this thesis. Breast cancer images in the mammogram modality have a lot of non-cancerous structures that are similar to cancer, which makes them especially difficult to work with. Only the cancerous part of the image is relevant, so the techniques must learn to recognize cancer in noisy mammograms and extract features from that cancer to classify or retrieve similar images. There are also many types or classes of cancer with different characteristics over which the system must work. Mammograms come in sets of four, two images of each breast, which enables comparison of the left and right breast images to help determine relevant features and remove irrelevant features. Image feature comparisons are used to create a similarity function that works well in the high-dimensional space of image features. The similarity function is learned on an underlying clustering and then integrated to produce an agglomeration that is relevant to the images. This technique diagnoses breast cancer more accurately than commercial systems and other published results. In order to collect new data and capture the medical diagnosis used to create and improve these methods, as well as develop relevant feedback, an innovative image retrieval, diagnosis capture, and multiple image viewing tool is presented to fulfill the needs of radiologists. Additionally, retrieval and classification of prostate cancer data is improved using new high-dimensional techniques like dimensionally-limited distance functions and dimensional choice.