CancerSeg-XA: Enhanced Breast Cancer Histopathology Segmentation Using Xception Backbone with Attention Mechanisms

Abstract

Breast cancer remains a formidable health challenge requiring advanced computational tools for accurate diagnosis and treatment planning. This study hypothesizes that modifications to the DeepLabV3+ architecture, such as incorporating an attention layer and replacing the ResNet50 backbone with Xception, can significantly enhance segmentation accuracy and model stability for breast cancer histopathological images. To test this hypothesis, we evaluated the performance of the original DeepLabV3+ and three modified versions for semantic segmentation using the “Breast Cancer Semantic Segmentation” (BCSS) dataset, which provides pixel-wise annotations of breast cancer tissues. The proposed modifications include integrating an attention layer between the encoder and decoder (Model 1), replacing the ResNet50 backbone with an Xception backbone up to 'block5' (Model 2), and combining the Xception backbone with the attention layer (CancerSeg-XA). The models were implemented and trained in the Kaggle Notebook environment, and their performance was assessed based on training and validation accuracy. The results show that Model 1 improved the model stability and accuracy compared to DeepLabV3+, whereas Model 2 and CancerSeg-XA achieved significant accuracy improvements of 91.47% and 91.57%, respectively, over the baseline DeepLabV3+ accuracy of 85.7%. CancerSeg-XA demonstrated enhanced training stability, making it a promising approach for clinical application in breast cancer diagnosis and treatment.