A Transformer-Based Hybrid Framework for Breast Tumor Segmentation Using CAS-UNet and ViT

When breast cells grow out of control, it causes breast cancer and tumors to form. It greatly endangers women's lives around the world, making it necessary to find better ways to discover and sort out the disease. The research proposes a new system that uses traditional image preprocessing combined with advanced deep learning models-CAS-UNet and ViT. Initially, the images from the database are adjusted with contrast-limited Adaptive Histogram Equalization (CLAHE) to highlight fine details and contrast and then are normalized and resized to keep their dimensions uniform. Following preprocessing, the images are sent to CAS-UNet, a transformer-based architecture that includes attention mechanisms and texture enhancement to better handle both the spatial details and fine features of the image. After that, the regions are improved and confirmed by a ViT, which increases the ability to distinguish features and helps explain the outcomes. The method suggested here showed strong and reliable performance by accurately identifying breast cancer regions with an accuracy of 99.03%, recall of 99.22%, precision of 99.31 % and an F1-score of 99.19%. Experiments reveal that the proposed CAS-UNet-based method works better than other methods in preserving the shape of boundaries and achieving accurate segmentation, making it ideal for clinical work and large-scale automatic diagnosis.