Deep Learning Based Automated Whole Breast Ultrasound Analysis:Anatomical Layer Parsing And Cancer Detection

Breast cancer is a serious disease for women all over the world.It is the most common cancer and the leading cause of cancer death in women.The most effective approach to reduce the mortality rate is early detection of the disease.To this end,breast imaging plays a fundamental role in detecting breast cancer at early stages.Automated breast ultrasound(ABUS)is one of the most popular imaging modalities for the early detection and diagnosis of breast tumor.In ABUS system,a huge transducer automatically moves across or around the breast and stacks 2D slices to form a 3D volumetric image.The ABUS imaging is fast and reproducible thus may lead to a revolution in breast imaging and cancer detection and diagnosis.However,reading ABUS images is extremely time-consuming,and the relatively high false positive rate remains a concern.Therefore,the automatic and intelligent analysis in ABUS images is expected to assist clinicians in facilitating the identification of breast cancer lesion.Specifically,in this paper,we propose an ABUS automatic anatomical decomposition system(ABUS-AADS)to quantitatively measure the density of breast tissues,as well as an ABUS automatic tumor detection system(ABUS-ATDS)to accurately and effectively identify the location of breast lesion.We first address the issue of anatomical layer segmentation in ABUS images.To tackle the issue of ambiguous boundaries in ABUS images,we integrate the boundary regularization based Deep Boundary Supervision(DBS)with Convolutional EncoderDecoder Network(Conv EDNet),and propose to train the network using a Two-Stage Adaptive Domain Transfer(2DT).Experimental results demonstrate that our method provides more accurate segmentation than state-of-the-art methods do,and our segmentation results are comparable with the annotations from experienced physicians.In addition,we develop a 3D-Unet based algorithm for the lesion detection in ABUS images.To improve the detection rate,especially the detection rate for smallsized lesions,we propose a Densely Deep Supervision(DDS)to reinforce the gradient of back-propagation in the deep learning network,and use the state-of-the-art video classification model C3 D to accelerate the convergence of the proposed network.To address the issues of unbalanced training data and the relatively high false positive rate,we further employ DSC Loss to supervise the network and use the Fully Connected CRF to remove false positive.Experimental results show that our method achieves the detection accuracy of 95.27% and 3.4 false positives in per-volume.To the best of our knowledge,the developed system in this study is the first deep learning based ABUS image segmentation and detection system,which would be widely applied in clinical practices in the near future.