Research On Computer-aided Detection Of Masses In Digital Breast Tomosynthesis Based On Deep Learning

Breast cancer is the most common disease that endangers women's health,because of its high incidence of disease and caused widespread concern and panic.One of the most important ways to reduce the mortality rate of breast cancer is early diagnosis and treatment.However,most patients have been diagnosed as breast cancer in the middle or late stage,delaying the best period of treatment.In the early diagnosis of breast cancer,full-field digital mammography(FFDM)is one of the preferred techniques,but this two-dimensional image still has the problem of missed diagnosis due to the overlap of dense breast tissue and lesions.The development of digital breast tomosynthesis(DBT)technology has greatly improved the clarity of the symbolic display of X-ray breast lesions.This three-dimensional imaging method can reduce or eliminate tissue overlap and display the structural features of surrounding tissues more clearly.Therefore,DBT imaging technology improves the detection rate of lesions and reduces the rate of missed diagnosis.Since DBT images are three-dimensional tomographic images,a sequence of images usually contains dozens of tomographic images.The workload of radiologist has increased dozens of times with DBT.Therefore,it is urgent to develop an intelligent computer aided diagnosis(CAD)System for DBT images to help doctors reduce workload and improve diagnostic efficiency.The computer aided detection system of mass is the Key algorithm in the intelligent assistant diagnosis system.This paper mainly studies the computer aided detection system of mass in DBT slices based on deep learning.In previous studies,most researchers have focused on designing traditional machine learning algorithms that extract features manually.Recently years,deep learning algorithm has excellent performance in image recognition and machine vision,hence many researchers have applied deep learning algorithn to automatic detection of masses in DBT image and achieved many excellent results.However,in these studies,the deep learning algorithm has not been applied to the whole DBT mass automatic detection system.The pre-selection step of suspected mass areas in the system is still based on the traditional algorithm of manual feature extraction.There are some problems,such as the system structure is not integrated enough,and the operation efficiency is not high enough.In this paper,two kinds of algorithm of breast DBT detection system of mass based on deep learning are designed.The first is detection system based on deep convolutional network(DCNN).The second is based on Faster Regional Convolutional Neural Networks(Faster RCNN).The performance of the two algorithms is evaluated and compared on the same data set.Specific contents include:(1)Image preprocessing:image contrast enhancement,breast region detection and breast skin segmentation of breast DBT images.(2)DCNN-based automatic detection system for DBT masses:Firstly,three-dimensional radial gradient analysis is used to compute the three-dimensional DBT slices,and the three-dimensional radial gradient feature map of the three-dimensional DBT slices is obtained.Then three-dimensional adaptive threshold segmentation is performed on the three-dimensional radial feature map,which is divided into three levels,and the highest-level region is retained as the suspicious region of the mass.Then,a 25.6×25.6 mm box was used to intercept five consecutive ROI(region of interest)regions at the center of the suspected region of the mass and input the deep learning convolution neural network model to obtain the likelihood score of the mass in the suspected region(3)Computer-aided detection of mass in DBT based on Faster RCNN:A framework of Faster RCNN deep learning object detection for mass detection is built.These include:shared convolution network,regional proposal network(RPN),false positive classification network.In the process of detection,the continuous DBT three-dimensional tomographic image layers are input into the trained Faster RCNN model.After five-layer shared convolution network,RPN automatically extracts the features of the shared convolution network and transfers the suspected region of the mass to the classifier network in the form of a boundary box,which is used to obtain the mass likelihood score.Then,the bounding boxes are fused based on the overlap rate of each other,and the maximum score is taken as the final likelihood score of the masses in the region to realize mass detection.(4)Performance evaluation method and performance comparison:The two systems are cross-validated by 10 folds using identical data sets,and the performance is evaluated by free-response receiver operating characteristic curve(FROC).When the sensitivity of DCNN-based automatic detection system for DBT masses reaches 90%.there are 2.25 false positives per breast,and 0.76 false positives per breast when the sensitivity of Faster RCNN-based automatic detection system for DBT masses reaches 90%.Using resampling nonparametric method to test the FROC of the two algorithms,it was found that there was significant difference between the two algorithms(P<0.05).Faster RCNN-based mass detection system for DBT performed better.The main contribution of this study is to use Faster RCNN framework in the computer-aided detection of mass in DBT images for the first time and achieve excellent results.At the same time,it is the first time to compare the performance of DBT mass computer-aided detection system based on deep convolution neural network with that of DBT mass computer-aided detection system based on Faster RCNN.It provides a new research direction in the computer-aided detection of mass in DBT.