AMRM: Attention-based mask reconstruction module for multi-classification of breast cancer histopathological images

Abstract

Nowadays, breast cancer is a leading cause of cancer-related mortality among women globally. Approximately 10% to 15% of breast cancer patients fail to undergo timely screening, resulting in a missed opportunity for optimal treatment. Computer-aided diagnosis (CAD) systems have been used successfully in breast cancer diagnosis. Nevertheless, current systems have encountered difficulties in achieving a high degree of accuracy, with the majority of research efforts focusing on the binary classification that distinguishes benign from malignant. Different subtypes of breast cancer require different targeted therapeutic approaches. Therefore, the precise classification of the breast cancer subtype has a major impact on treatment decisions. To improve the accuracy of breast cancer multi-classification, a novel Attention-based Mask Reconstruction Module (AMRM) is proposed to improve the performance of the model. AMRM extracts features from breast cancer histopathological images through the attention module and then performs mask reconstruction to generate reconstructed features. These reconstructed features were used in a multi-classification task to accurately classify histopathological images of breast cancer. AMRM enables the network to effectively identify background and foreground in histopathological images, reduce background interference, improve adaptability to background changes, align the features extracted by the model with the pathologist's expectations, and improve classification accuracy. Results from experiments conducted on the BreakHis dataset show that the inclusion of AMRM resulted in a significant improvement in multi-classification accuracy for the AlexNet, VGG11, ResNet-50 and Data-efficient Image Transformer (DeiT) models, reaching 88.48%, 93.40%, 96.49% and 94.10% respectively. Compared to the baseline model, accuracy increased by 8.28%, 2.11%, 1.27% and 1.26% respectively, demonstrating a significant improvement.