Reliable Breast Cancer Detection from Ultrasound Images using Image Segmentation

Health 4.0 presents an intriguing vision for the healthcare sector. It effectively combines cutting-edge technologies such as Healthcare Cyber-Physical Systems, medical image analysis, and Deep Learning algorithms in order to improve the efficiency of the cancer diagnostic process. Breast cancer is arguably one of the primary factors contributing to mortality rates among women worldwide. In this paper, a method of identifying breast tumors from ultrasound images using image segmentation and a reliable version of the U-NET model is presented. To enhance its robustness, several regularization methods are used, such as L2 norm and Dropout layers. The model is trained on a dataset containing 780 images that were augmented to 3900 images using contrast adjustment. Additionally, the model was validated on an external dataset containing 163 ultrasound images. The model achieved an accuracy of 98.23% and a dice score of 93.92% when evaluated on the training dataset. Regarding the external dataset, the model achieved an accuracy of 94.28% and a dice score of 90.15%.