Efficient Deep Networks for Tumor Segmentation in Breast Ultrasound Imaging

Breast cancer is regarded as one of the major causes of cancer-related deaths among women in today’s world. Accurate detection at an early stage plays a crucial role in the improvement of survival rates. In such a scenario, ultrasound imaging has emerged as a safe and cost-effective screening modality, particularly in resource-constrained settings. However, automated tumor segmentation in ultrasound images is a challenging task due to low contrast, speckle noise and heterogeneous lesion boundaries. This study presents a comparative evaluation of two lightweight deep learning architectures, MobileNetV2–UNet and MobileNetV2–DeepLabV3, for automated breast tumor segmentation using the BUSI dataset. The study focuses exclusively on benign and malignant categories. Each model is assessed based on segmentation accuracy (Dice coefficient, IoU, precision, recall) and computational efficiency (inference time and model size). Experimental results demonstrate that MobileNetV2–UNet achieved 79.50% IoU, 86.14% Dice coefficient and an overall accuracy of 97.95%, with an inference speed of 6.54 FPS. In comparison, MobileNetV2–DeepLabV3 attained 78.61% IoU, 86.06% Dice coefficient and 98.02% accuracy, but with a slower inference speed of 2.02 FPS. Additionally, inference benchmarking demonstrates that MobileNetV2–UNet delivers 3× faster inference speed, making it better suited for edge deployment in mobile health screening applications. Both models demonstrated robust tumor localization performance and edge deployability. This work contributes an efficient, scalable solution for automated breast lesion segmentation, relevant for deployment in clinical settings with limited computational resources.