Breast Ultrasound CAD System Based on Efficient Tumour Segmentation Network and Transfer-Learned Features
Nadeem Zaidkilani, Mohamed Abdel‐Nasser, Miguel Ángel García, Domènec Puig · 2022
Breast cancer is the second most common type of cancer worldwide after lung cancer and the leading cause of cancer death among women. Over the past few decades, computer-assisted diagnostic (CAD) systems have been implemented to assist physicians. This paper introduces a CAD system to segment tumours in breast ultrasound (BUS) images and classifies them as benign or malignant. The CAD system has two stages: segmentation and classification. In the segmentation stage, we have developed an encoder-decoder network based on different backbones with various loss functions to segment the tumours. We have fine-tuned the MobileNetv2 network in the classification stage to classify the segmented tumours as benign or malignant. Our experiments demonstrate that WideResNet with BCE and Dice loss function outperforms and yields the best tumour segmentation results with a Dice score of 77.32%. The CAD system achieves a classification accuracy of 86%.