Impact of Data Augmentation on U-Net Architectures for Breast Tumour Segmentation in Ultrasound Imaging

Breast ultrasound is one of the most widely used imaging modalities for detecting breast tumours. However, its analysis remains challenging due to lesion variability and demands on medical resources. Therefore, automatic segmentation is crucial. This paper evaluates the performance of the U-Net and Attention U-Net architectures in segmenting benign and malignant breast tumours in ultrasound images, emphasising the importance of data augmentation. An accuracy above 91% was consistently achieved across all experiments. Data augmentation improved precision in malignant cases, from 72.67% to 88.16% (U-Net) and from 79.51% to 89.78% (Attention U-Net). However, it reduced recall, indicating a more conservative segmentation. In contrast, benign cases demonstrated simultaneous improvements in precision and recall with data augmentation, particularly with the Attention U-Net. These results highlight the critical relationship between recall and precision in malignant tumour segmentation and reinforce the value of combining deep learning with data augmentation to increase the robustness and clinical applicability of breast ultrasound analysis.