Deep Learning-Based Approach for Breast Lesion Segmentation in Ultrasound Images

Abstract

This thesis explores deep learning-based approaches for automated breast lesion segmentation in Ultrasound (US) images, aiming to enhance the accuracy and efficiency of computer-aided diagnosis (CAD) systems. Breast cancer diagnosis often faces challenges due to the variability in lesion morphology, low contrast in US images, and limited annotated datasets. To address these issues, this study investigates state-of-the-art segmentation architectures, including U-Net variants, transformer-based models, and hybrid CNN-transformer frameworks, alongside advanced preprocessing and augmentation techniques. The proposed methodology evaluates baseline and novel models using metrics such as Dice score and Intersection-over-Union (IoU), with a focus on improving robustness to dataset imbalances and image complexity. Experimental results demonstrate the effectiveness of deep learning in achieving precise lesion delineation, offering potential clinical utility for early and non-invasive breast cancer and other tumor detection.