CBIR for mammograms using medical image similarity

One fundamental problem remains in the area of medical image analysis and retrieval: how to measure radiologist's perception of similarity between two images. This paper develops a similarity function that is learned from medical annotations and built upon extracted medical features in order to capture the perception of similarity between images with cancer. The technique first extracts high-level medical features from the images to determine a local contextual similarity, but these are unordered and unregistered from one image to the next. Second, the feature sets of the images are fed into the learned similarity function to determine the overall similarity for retrieval. This technique avoids arbitrary spatial constraints and is robust in the presence of noise, outliers, and imaging artifacts. We demonstrate that utilizing unordered and noisy higher-level cancer detection features is both possible and productive in measuring image similarity and developing CBIR techniques.