Enhanced Breast Cancer Detection in Ultrasound Imaging Using an Attention Based U-Net++ Architecture for Improved Tumor Segmentation

Breast cancer, the most commonly occurring cancer among women and a leading reason of cancer mortality worldwide, necessitates early detection to reduce mortality rates. Breast Ultrasound (BUS) imaging, coupled with Computer Aided Diagnosis (CAD) systems, has emerged as a crucial, non-invasive, non-radioactive and low cost method for early breast cancer detection. Its suitability for large-scale screening and diagnosis, particularly in low-resource settings, further underscores its importance. However, the automatic segmentation of tumors in BUS images remains disputing due to the lower quality of the image, characterized by speckle noise, low contrast, feeble boundaries and artifacts. A histogram equalization was used as a pre-processing model to overcome such an issue. Additionally, the significant variation in tumor shape, size and echo strength across patients complicates the application of conventional segmentation methods that rely on strong priors to object features. This work proposes a novel attention-based U-Net++ architecture designed to address these challenges and achieve highly accurate breast cancer segmentation in BUS images. This model integrates an attention mechanism to boost the network’s ability to attention to pertinent image regions, improving tumours’ delineation despite noise and artifacts. The architecture incorporates the U-Net++’s strength in handling biomedical image segmentation tasks, incorporating skip connections to preserve spatial information and attention gates to filter and refine features at multiple scales. Through extensive experimentation and evaluation using Python, the attention-based U-Net++ demonstrates superior performance in segmenting tumors in BUS images, outperforming conventional models in terms of accuracy and robustness.