Theses and Dissertations Collection

Deep learning-based computer-aided diagnosis systems have achieved unprecedented performance in breast cancer detection. However, most approaches are computationally intensive, and impede their broad dissemination. In this study, we propose an efficient and light-weighted multitask learning framework to simultaneously classify and segment breast tumor. Pretrained MobileNetV1 is employed as the multitask network backbone, followed by 2 branches for classification and segmentation respectively. The segmentation branch utilizes the Link-Net as decoder. The proposed approaches are evaluated using a dataset with 864 B-model breast ultrasound images. Extensive experiments demonstrate that the proposed multitask learning network not only improves the classification and segmentation accuracy, but also keeps the low latency and efficiency properties. The network achieves 86.6% Dice’s coefficient and 79% Intersection Over Union for segmentation, while 93.85 % Accuracy, 94.44% Sensitivity and 93.42 Specificity for classification. The trained network has the size of around 60 MB in Keras H5 format, and 20 MB after converting to Tensorflow Lite format. Lastly, we develop and build a mobile application in Android Studio. It launches the trained multitask learning network to do real-time breast tumor detection, with average inference time cost for classification and segmentation together being around 300 milliseconds.