Efficient-Net ASPP Deep Network for Malignant Ultrasound Breast Cancer Segmentation

Abstract

Worldwide, breast cancer is one of the main causes of mortality among women. Only through early recognition of symptoms is it possible to limit the incidence of premature deaths. Utilizing standard deep-learning semantic segmentation, tumors are automatically segmented. It requires minimal human contact. The standard methods are less efficient and precise in accuracy. In this paper, we propose an Efficient-Net based Atrous Spatial Pyramid Pooling (Efficient-Net ASPP) model for efficient and improved segmentation. The proposed model is comprised of an EfficientNetV2B3 backbone with dilated convolutional, skip connection, max pooling, and an ASSP block. A publicly available Malignant Breast Cancer Ultrasound (MBCU) subset image dataset was investigated in our study. Several baseline segmentation methods were compared with our proposed method. The quantitative benchmark analysis shows the proposed model for segmenting MBCU images achieved 70% intersection over union (IOU) and a 62% score on the Dice Similarity Coefficient (DSC) metrics. The proposed network improves MBCU image segmentation by 14% on DSC and 18% on IOU when compared to baseline models.