Wavelet Convolution Neural Network for Classification of Spiculated Findings in Mammograms

Abstract

The subject of this paper is computer-aided recognition of spiculated findings in low-contrast noisy mammograms, such as architectural distortions and spiculated masses. The issue of computer-aided detection still remains unresolved, especially for architectural distortions. The methodology applied was based on wavelet convolution neural network. The originality of the proposed method lies in the way of input image creation. The input images were created as the maximum value maps based on three wavelet decomposition subbands (HL,LH,HH), each describing local details in the original image. Moreover, two types of convolution neural network architecture were optimized and empirically verified. The experimental study was conducted on the basis of 1585 regions of interest (512 x 512 pixels) taken from the Curated Breast Imaging Subset of Digital Database for Screening Mammography (CBIS-DDSM), containing both normal (1191) and abnormal (406) breast tissue images including clinically confirmed architectural distortions (141) and spiculated masses (265). With the use of wavelet convolutional neural network with a reverse bioorthogonal wavelet, the recognition accuracy of both types of pathologies reached over 87%, whereas the recognition accuracy for architectural distortions was 85% and for spiculated masses - 88%.