A microcalcification cluster detection method based on deep learning and multi-scale feature fusion

Abstract

To accurately identify microcalcification clusters (MCs) in x-ray images to detect breast cancer earlier, a new MC target detection method which combines a deep fine-grained cascade-enhanced network (FCE-Net) and a multi-scale feature fusion algorithm (MFF) is proposed. First, a deep convolutional neural network model FCE-Net is established to extract characteristics of MCs. FCE-Net constructs a convolutional subnetwork within the convolutional module to enhance the multi-branch structure and thus improves the detailed feature extraction performance of MCs. Then, a multi-level, fine-grained convolutional map is generated and a multi-scale feature fusion algorithm (MFF) is constructed. In order to solve the problem of fine-grained target semantic information enhancement, MFF up-samples the basic feature map twice and then laterally connects to the previous layer. Then, through the parallel regression calculation and feature classification, the confidence and regional coordinates of MC objects are obtained. Finally, the target area is classified and bounding box adjustment is performed in the merged layer of the region of interest. We perform the proposed method on the MIAS breast cancer data set. By training and testing, the accuracy of detection is 97.16% under the N5 parameter settings. The overall model improves the detection efficiency of small targets by 5–10%.