Breast Cancer Segmentation From Ultrasound Images Using Multiscale Cascaded Convolution With Residual Attention-Based Double Decoder Network

Accurate segmentation of breast cancer (BC) in ultrasound images is a complicated task due to the variable nature of ultrasound images. Recently, many techniques are suggested to accurately segment BC using ultrasound imaging and deep learning. A multiscale cascaded convolution with residual attention-based double decoder network for BC segmentation is presented in this study. A multiscale cascaded convolution operation-based encoder path is designed to overcome the problem of a single scale feature learning process. The proposed multiscale convolution operation helps to extract the diverse semantic spatial features. In the segmentation process, a residual attention-based double decoder network is proposed. The proposed attention mechanism is implemented to take out the more prominent features from the tumor region and to suppress the other information that can mislead the segmentation model during training. The double decoding mechanisms are introduced to capture the highly diverse spatial features that are learned in the encoder path. For experimental purposes two publically available ultrasound image datasets namely BUSI and UDIAT are utilized. The proposed U-shaped multiscale cascaded convolution with residual attention-based double decoder network achieved the segmentation dice of 91.38% with 81.67% of the Jaccard index, 94.43% precision, and 87.76% recall score on the UDIAT dataset. A dice score of 90.55% is recorded with a Jaccard of 80.87%, 93.53 % precision score, and 88.46% recall score on the BUSI dataset. The results of the multiscale cascaded convolution with residual attention-based double decoder network validated that it can effectively be used for breast cancer segmentation tasks.