Breast cancer has the highest incidence among Chinese female,which is increasing year by year.Compared with other cancer diseases,breast cancer has a higher cure rate if it is detected at an early stage and treated effectively.Therefore,early detection research has important clinical significance.In recent years,automated breast ultrasound(ABUS)has played an important role in the screening and diagnosis of breast tumors.However,manual reading of images brings a huge workload to doctors and has a certain subjectivity.Therefore,developing an accurate and efficient computer-aided ABUS tumor detection system is of great significance for clinical auxiliary diagnosis.This paper conducts a study on tumor detection methods in ABUS images based on deep learning.The main work and innovations of this paper are summarized as follows:(1)In order to alleviate the problem of false detection and missed detection of the tumors caused by excessive noise and the similarity between the tumors and the normal hypoechoic tissue in ABUS images,a tumor detection method based on the BayesianYOLOv4 network is proposed.In this method,some tricks are optimized on YOLOv4 according to the characteristics of ABUS images.Aiming at reducing the epistemic uncertainty of the model,the uncertainty information is fused into the optimized model,which reduces the false positive detections caused by the lack of awareness of the ABUS image and improves the tumor cognitive ability of the model.In addition,in order to highlight the foreground region of the tumor,a segmentation branch is proposed to assist the tumor detection without additional labeling cost,and a segmentation loss based on a2 D Gaussian mask is constructed.Experimental results prove that compared with the original YOLOv4 network,the multi-task Bayesian-YOLOv4 model incorporating uncertainty information can effectively alleviate the problem of excessive false positives and improve the tumor detection rate.(2)Aiming at addressing the high cost of ABUS image labeling and the imbalance of image foreground and background,a semi-supervised learning network model SSLEfficient Det is proposed,which combines self-training and consistency regularization.In the interest of making the model learn more reliable information,a pseudo-label strategy based on confidence filtering is designed.A consistency regularization algorithm based on data augmentations is deployed to construct an unsupervised loss,so that the model can obtain better generalization performance under the combined effect of the supervised loss and the unsupervised loss.In addition,in order to tackle the problem of unbalanced foreground and background in ABUS images,a Copy-Paste strategy based on Gaussian blur is designed,which increases the number and diversity of foreground samples.Experimental results demonstrate that compared with the fully supervised model,the proposed semi-supervised learning model SSL-Efficient Det effectively utilizes unlabeled data and achieves better detection performance.(3)To resolve the false detection and missed detection of ABUS images caused by the overconfidence predictions of a single network model in the deep-learning based method,a deep mutual learning DML-YOLOX model based on Exploration Loss and Consistency Loss is proposed.In this method,in order to discover the global optimal solution of the model,a dual network model is designed,which constructs the Exploration Loss and the Consistency Loss to restrain each other.The former encourages models to explore and learn different feature representations in the feature extraction stage of the network,while the latter constrains each model to have consistent output representations for the same input in the prediction stage of the network.According to the ellipse-like characteristics of the tumors in ABUS images,a data augmentation method based on ellipse rotation is designed to enhance the diversity of tumor samples and the generalization ability of the model.In order to reduce the uncertainty of position regression caused by large angle rotation,a rotation uncertainty loss is designed.Experimental results show that compared with the single network model,the tumor detection performance based on DML-YOLOX network is effectively improved. |