Cameron Kyle-Davidson, University of York
When radiologists evaluate mammograms images from the left and right breasts are shown concurrently. Radiologists remain capable of detecting abnormalities even up to three years prior to onset of cancer, and when said mammograms are presented rapidly. However, if the normal mammogram contralateral to the abnormal mammogram is replaced by that of another woman, this ability suffers a performance decrease. Evidently, a global signal that signals abnormality exists and is dependent on both mammograms. We investigated whether the effect also appears in a pre-trained neural network mammography model. Further we explored the effect of bilateral differences by developing and training a neural network model which can reliably detect whether a set of mammograms is composed of images taken from the same woman, or two different women. Detection of bilateral asymmetry remains even when mammograms are balanced by size and age; indicative that a “symmetry signal” exists and is relevant for breast cancer detection. We pilot off-site cloud GPU resources for both training and inference of the neural networks, which would have been intractable on our local hardware. In addition, we develop a semi-autonomous mammography dataset cleaning pipeline that can take advantage of high-cpu count cloud machines; through multithreaded image processing.