Research On Tumor Detection Method In ABVS Images Based On Deep Learning

Breast cancer is one of the leading causes for cancer death in women.Its incidence rate is increasing year by year and shows a younger trend.Recently,automated breast volume scanner(ABVS)has been applied in clinical practice.The application of ABVS in the detection and diagnosis of breast tumor has drawn more attention.As ABVS realizes the acquisition of continuous transverse panel slice,screening ABVS images is a time-consuming work for radiologist.Besides,the screening result based on subjective analysis has certain individual differences.Computer-aided diagnosis system based on image processing and pattern recognition,can reduce the cost of diagnosis,improve the efficiency and objectivity,which has important research value.At present,tumor detection based on ABVS images bears the following problems.(1)The noise produced by the uneven application of couplant and the fatty tissue show hypoechoic characteristics.Moreover,the shape and contrast of tumor in different cases are quite different.These issues make the correct tumor detection very challenging.(2)Existing tumor detection methods treat each ABVS slice as an independent individual,which ignores the spatial continuity in ABVS images.Thus,tumor position in adjacent slices may have large difference and false positive detections are increased.(3)Tumor labeling of ABVS images should be completed by experienced radiologist.The heavy labeling work increases the burden of radiologist.In order to solve the above problems,tumor detection methods in ABVS images based on deep learning are studied in this thesis.The main work and innovation are summarized as follows.(1)Tumor regions are difficult to identify and false positive regions are prone to be detected.To solve these problems,a single slice tumor detection method based on Improved-YOLOv3 and post-processing algorithm is proposed.In order to improve the network detection performance for tumor,a mask generation branch is constructed,which has rich semantic information and high-resolution spatial information.Then a multi-task learning network Improved-YOLOv3 is designed,which is jointly trained by the detection loss and the tumor segmentation loss.In order to improve the continuity of tumor location in adjacent slices and remove false positive regions,a post-processing algorithm is designed,which includes re-scoring detection results,three-dimensional volume generation and false positive removal by fusion of continuous slice detection results.Experimental results show that,compared with YOLOv3 network,the multi-task learning network Improved-YOLOv3 combined with post-processing algorithm improves the tumor detection accuracy and reduces the false positive rate.(2)Single slice tumor detection method ignores inter-slice information.Aiming at this problem,an SC-BLSTM-YOLOv3 network is proposed to realize end-to-end continuous slice tumor detection.To integrate the inter-slice information in tumor detection,bidirectional long short term memory(BLSTM)structure is embedded in YOLOv3 detection network.To solve the problem of semantic information gap in the feature pyramid of YOLOv3,the spatial-channel(SC)attention mechanism is combined with the feature pyramid.Experimental results show that the SC-BLSTM-YOLOv3 network,integrating inter-slice information and attention mechanism,has better performance for tumor detection in ABVS images.(3)The labeled ABVS dataset is difficult to obtain.Aiming at this problem,a semisupervised detection network SSL-YOLOv3 based on self training model is proposed.To obtain high quality pseudo labels,a pseudo label update strategy based on validation set is designed.In order to enhance the robustness of the detection model,strong data augmentation method suitable for target detection task is adopted.Experimental results show that the semi-supervised detection network can make full use of large unlabeled data and obtains better detection performance than the model trained only with labeled dataset.