Image Differencing Approaches to Medical Image Classification

Image differencing is normally done by subtracting the low-level features like intensity in images that are already aligned. This paper extracts high-level features in order to learn an effective image differencing method for the diagnosis of breast cancer. However, this produces sets of features that are both spatial and unordered. Learning techniques are challenging when examples are sets of features that lack any sort of meaningful ordering and where spatial relationships are important. We demonstrate a technique that avoids arbitrary spatial constraints and is robust in the presence of noise, outliers, and imaging artifacts, while outperforming even commercial products in the diagnosis of breast cancer images. First, the landmarks are found and ranked, and then the top candidates are sorted into a point set. Second, the point sets of the two images are then differenced through a cluster comparison. A technique that radiologists use to diagnose breast cancer involves finding potentially cancerous sites in the mammograms and then comparing the left and right breasts to reduce the effect of false positives and to produce a diagnosis. The symmetry of the human body is utilized to increase the accuracy of the diagnosis. We emulate this technique in an attempt to understand and eventually capture the diagnosis of the radiologist. The image differencing with clustered comparison process determines the presence of cancer 80% of the time, outperforming all other systems, thus making it a strong classifier which should significantly improve systems for helping radiologists diagnose breast cancer images. The results compare favorably with the state of the art in both academic and commercial approaches, achieving a 9% overall improvement over the best academic approach and a 26% improvement on non-cancerous cases over the best commercial approach while maintaining the same accuracy on cancerous cases.