Asymmetry Analysis with Sparse Autoencoder in Mammography
Dongxiao Yang, Ying Wang, Zhicheng Jiao · 2016
With the development of medical image processing methods, computer-aided detection systems sprang up, which could assist the diagnosis of radiologists, reducing fatigue, and greatly improve the accuracy and efficiency of early detection. However, most researches concentrated on mass detection methods by unilateral mammography, ignoring the relationship between left and right mammograms that the bilateral breast tissues were symmetric by physiology. However, radiologists often preferred to take images of both sides into consideration to judge if there were lesions. In this paper, we proposed an asymmetric analysis framework which employed bilateral mammograms together. Generally speaking, an improved bilateral region matching method that control points extracted automatically based on shape context was used to align the corresponding regions of bilateral mammograms. Meanwhile, sparse autoencoder was introduced into asymmetric analysis to learn a network between normal and its corresponding normal regions. For giving one side of bilateral regions, it would output a reconstruction vector of another form through the trained network. A similarity was computed between the output and the corresponding region of input, therefore a reasonable result was given.