Attention-Enhanced Multi-Task U-Net for Breast Cancer Detection and Automated Risk Assessment

Abstract

Early detection of breast cancer through mammograms is a difficult process since mammograms have a poor contrast ratio and contain noise. Though U-Net architectures have provided good results for image processing in medicine, their performance is greatly affected by poor feature extraction and lack of integration with real-time diagnostics. This paper puts forward an attention-assisted multi-task U-Net architecture which aids in detecting breast cancer lesions along with their real-time risk estimation. Attention has been applied within the architecture to make it learn features more accurately and localize suspicious regions, whereas the use of a multi-task approach allows the model to detect lesions as well as assess their risk. Moreover, a pipeline has been devised to facilitate real-time interaction with the system using mammograms. The developed model is trained and tested using benchmark mammography datasets and attains a classification accuracy of 97.5%, along with higher robustness than existing CNN-based models. The experimental findings demonstrate greater reliability in terms of cancer detection, especially during its early stages. The designed framework offers a practical and affordable approach to breast cancer screening by connecting detection using deep learning algorithms with real-time decision-making.