A New Triplet Convolutional Neural Network for Classification of Lesions on Mammograms
Medjeded Merati, Saïd Mahmoudi, Abdelkader Chenine, Mohamed Amine Chikh · Revue d intelligence artificielle · 2019
Mammography provides a useful tool for breast cancer detection.However, many doctors have difficulty in making the right decision based on mammograms.This paper aims to set up a deep learning (DL) architecture that can effectively differentiate between benign and malignant tumors.Specifically, a new triplet convolutional neural network (CNN) was established with three subnetworks, each of which contains a succession of layer blocks.Each block consists of two convolutional layers, a dropout layer and a max-pooling layer.During operation, the region of interest (ROI) extracted from the mammogram is imputed to the first subnetwork, and processed by the Canny filter.The filtered results become the input of the second subnetwork, while the third subnetwork takes the whole image as input.To verify the effectiveness of our architecture, a set of 500 images from 301 patients was extracted from the DDSM database and augmented to 4,000 images, and divided into a training set (80 %) and a testing set (20 %).The results show that our architecture achieved an accuracy of 93.13 %, a sensitivity of 96 % and a specificity of 90.25 %.This research provides a desirable way to identify breast cancer based on mammography.