Research On Digital Breast Tomosynthesis Mass Detection And Segmentation Based On Convolutional Neural Network

Breast cancer is one of the common malignant tumors in women.It has drawn wide attention due to high morbidity and high mortality.The experimental and clinical study has shown that early detection,diagnosis,and treatment will greatly decrease mortality rate.Mass is the first diagnosed symptom of 80% patients with breast cancer.So it is very important for diagnosing breast to check mass accurately.Digital breast tomosynthesis(DBT)can get information from image in each depth of breast to reduce influence of overlying tissue in diagnosis.Although DBT has excellent properties,it is not popularized domestically.The burden of radiograph reading and the time of diagnosis are influenced which can be one factor.by the number of DBT images.DBT has larger number of images than FFDM because of its imaging model.Computer Aided Diagnosis(CAD)for DBT is a feasible solution to improve the efficiency of diagnosing breast cancer.The algorithm of detection and segmentation of mass is one of core sections of CAD system.Detection of mass can provide interested areas which help doctor determine the location of the mass.Direct effect of segmentation of mass is providing contour of the mass for doctor.It can be convenient for the research which need more accurate information of the mass,such as identification of benign mass and malignant mass and the determination of the molecular classification.In this dissertation,convolutional neural networks was used to study the detection and segmentation methods of mass.In order to explore whether the detection performance and segmentation accuracy can be improved by using the three-dimensional information of DBT images,the method based on twodimensional convolutional neural networks and the method based on three-dimensional convolutional neural networks were compared.Specific contents include:(1)Research on mass detection and segmentation based on two-dimensional convolutional neural networks: A single DBT image was detected by Faster RCNN(Faster Region based Convolutional Neural Networks).Detection results of single image were fused and filtered to obtain predicted final detection regions.When the sensitivity reaches 90%,there are 3.14 false positives per breast.Spatial Fuzzy C-Means(SFCM)segmented mass in final detection region.After three steps,detection and segmentation results of DBT are obtained.Pixel accuracy of segmentation is 0.73.Mask RCNN(Mask Region-based Convolutional Neural Networks)was used to detect and segment the mass on single DBT image.Segmentation and detection results in one single were fused and filtered to get the final detection and segmentation results.When the sensitivity reaches 90%,there are 2.02 false positives per breast.Pixel accuracy of segmentation is 0.54.(2)Research on mass detection and segmentation based on three-dimensional convolutional neural networks was followed by next steps: Preprocessing the data.It includes extracting the breast region and clipping images;3D Mask RCNN was designed for detecting and segmenting the processed data;Designing the method of filtering and fusing results for detection and segmentation results to get final results.When the sensitivity reaches 90%,there are 1.23 false positives per breast.Pixel accuracy of segmentation is 0.81.In the end,this research was compared with the study based on two-dimensional convolutional neural networks.The comparison results show that the method based on three-dimensional convolutional neural networks is superior to the method based on two-dimensional convolutional neural networks in detection performance and segmentation accuracy.This research has potential implications for the future development of DBT CAD systems.