Real‐time automated segmentation of breast lesions using CNN‐based deep learning paradigm: Investigation on mammogram and ultrasound

Abstract

AbstractThe existing studies involving single imaging modalities (i.e., mammogram (MG) or ultrasound (US)) to detect breast lesions have demonstrated limited clinical application because radiologists rarely interpret an MG without a corresponding US and vice‐versa. Thus, this article aims to develop a Computer Aided Segmentation (CAS) system for detecting breast lesions in both MG and US. A customized convolutional neural network (CNN) is adopted for this purpose. A new real‐time bi‐modal database of MG and US is used for dual‐modality evaluation. Twelve performance measures, five shape measurements, area under receiver operating characteristics (ROC), and paired T‐test are used to assess the performance of proposed CAS system. A Dice Similarity Coefficient (DSC) of 0.64 (for MG) and 0.77 (for US) and Jaccard Index (JI) of 0.53 (for MG) and 0.64 (for US) indicate that the US can be used as an adjunct technique to MG in the segmenting breast lesions.