Abstract
In recent years, progress in breast cancer mortality reduction has slowed. The mortality rate was increasing and Breast cancer became the first cause of death for women. The early diagnosis is crucial in a treatment process as it may avoid the complications as well as the heavy treatment of the pathology. For this purpose, a lot of CAD systems were established. However, to provide more precise results, the later still needs to improve by incorporating new techniques. In this paper, we present a deep learning framework built on the U-Net architecture. A MobileNetV2 and a VGG16 model encoder have been used to handle the semantic segmentation of a biomedical image effectively. This approach is based on the integration of these pre-trained models with the UNet and having an efficient network architecture. By transfer learning, these CNNs are fine-tuned to segment Breast Ultrasound images in normal and tumoral pixels. An extensive experiment of our proposed architecture has been done using Breast Ultrasound Dataset B. Quantitative metrics for evaluation of segmentation results including Dice coefficient, Precision, Recall, and, all reached over 80%, which proves that the method proposed has the capacity to distinguish functional tissues in breast ultrasound images. Thus, our proposed method might have the potential to provide the segmentation necessary to assist the clinical diagnosis of breast cancer and improve imaging in other modes in medical ultrasound.