A deep learning approach for breast cancer diagnosis in ultrasound images

Abstract

Breast cancer is a leading cause of morbidity and mortality among women in Sub-Saharan Africa. However, the most popular breast lesion screening modality, ultrasound, yields noisy images prone to subjective radiological interpretation. Machine learning has the potential to solve this challenge. However, prior approaches only focused on the lesion in the image, and datasets used were not representative in the sub-Saharan context. In our proposed work, we developed a morphology-aware deep learning model that di↵erentiates between suspicious and non-suspicious breast lesions using breast ultrasound images acquired from Sub-Saharan Africa. We used three datasets in this work. We acquired two datasets from Cairo university and Breast and Axilla websites, and the third dataset was acquired from ECUREI, yielding a combined dataset. We used YOLOV4-tiny as our primary lesion detection algorithm. The YOLOV4-tiny model was trained on 1,033 images from public domains and 144 images from ECUREI and tested on 83 images from ECUREI. Our best model reveals a sensitivity and specificity of 88% and 89%, respectively, on test data. A comparison of our model with other state-of-the-art object detection and classification algorithms of SVM, KNN, VGG16, and EcientDet shows superior performance. Our model, compared with several object detection deep learning models from similar studies, illustrates competitive performance. Our approach is a promising approach towards the automation of breast lesion detection from breast ultrasound images obtained from Sub-Saharan Africa.