Comparing Different Deep-Learning Models for Classifying Masses in Ultrasound Images

Breast cancer is still the predominant type of cancer that affects women worldwide. Artificial intelligence (AI) developers are making great efforts in developing automated computer-aided detection and diagnosis (CAD) systems for breast cancer prognosis and classification. There are currently no official guidelines recommending the use of AI with ultrasound in clinical practice, and further research is needed to investigate more advanced approaches and demonstrate their utility. In this paper, we use one of the state-of-the-arts segmentation model based on deep learning, as well as compare different recent pre-trained models for the purpose of segmenting and distinguishing between benign and malignant masses in breast ultrasound images. The results showed that among the compared models, the most accurate model was the EfficientNetB7 model which achieved 88% on the segmented input images, that were segmented using the ResUNet model with 0.8932 Dice score and 0.8572 Iou score.