Deep Neural Network Method For Lesion Detection And Diagnosis Based On Digital Breast Tomosynthesis

Breast cancer is the most common malignant tumor in women.Since its high incidence,it has received wide attention from researchers.Studies have shown that screening can prevent breast cancer,find the lesion and make a timely diagnosis,thereby greatly reducing the mortality rate of breast cancer.Digital Breast Tomosynthesis uses X-ray tube rotating exposure within a certain range to reconstruct a series of two-dimensional images that are parallel to the detector.The images taken by DBT technology reduce the tissue overlap in the previous X-ray images,and the local feature imaging is clearer.However,due to the large amount of images in the shooting sequence,the burden of doctors in reading images has increased.Therefore,it is necessary to develop a computer aided diagnosis system based on DBT images to reduce the burden of doctors on reading pictures and improve diagnosis efficiency.At present,traditional machine learning methods are mostly used in the field of intelligent diagnosis of DBT imaging lesions.The steps are cumbersome,and it is necessary to manually mark the region of interest and establish a classifier.With the popularity of deep learning algorithms,scholars began to apply deep learning to detect lesions and diagnosis of them.However,it only uses deep learning to achieve a single detection or diagnosis task,and it still needs to manually determine the region of interest from the image,which does not achieve complete automation,and there are problems in software application.With the goal of developing an auxiliary diagnosis system,this research has carried out research on auxiliary diagnosis methods for microcalcification cluster detection,mass detection,segmentation and its diagnosis,as follows:(1)Research on the detection of microcalcification clustersAiming to position fast and accurately of microcalcification clusters in DBT images.In this paper,the target detection network Faster RCNN model combines tomographic screening and fusion strategies to realize the intelligent detection of DBT image microcalcification clusters.The results of the study show that when the sensitivity of the model established by this method is 0.95,the average false positive result per view is 0.15.(2)Research on mass detection,segmentation and diagnosisThe research on the intelligent diagnosis method of DBT image mass is an important part of this research.This paper designs a multi-task learning method for mass detection and diagnosis of the Efficient Det detection network on two-dimensional DBT images based on different backbone networks,and obtains the auxiliary diagnosis results of the sequence through screening and fusion.Among them,the Efiicient Det-D0 model has a better overall performance on the test set.When the sensitivity is 90%,an average of 0.24 false positive results per image are detected.Research based on the Mask Scoring-RCNN model uses the Res Net101 backbone network to construct a multi-task model for two-dimensional DBT image mass detection,segmentation and diagnosis.Its pixel accuracy reaches 0.63,and the detection accuracy is similar to Efficient Det D1.By using ScoringRCNN,0.42 false positive results are detected per view.The two types of models have not obtained obvious advantages in the diagnosis of benign and malignant in detection results.Therefore,this study,based on DBT three-dimensional mass data,using traditional machine learning methods and 3DRes Net to carry out benign and malignant diagnosis research,and compare the two types of experimental programs and results.Using traditional machine learning methods,46 features were recursively selected from the extracted radiomics features.GBDT was used as the classifier to achieve an AUC of 0.91,while 3D-Res Net with strategy of multi-stage transfer learning achieved an AUC value of 0.86.(3)Development of a prototype system for DBT imaging intelligent aided diagnosis.This research combines the characteristics of current image browsing software,thinking of software design and practical considerations,and develops intelligent aided diagnosis software.The Faster-RCNN(backbone network Res Net101)model is used to realize micro calcification detection,and the Mask Scoring-RCNN model is used to realize software mass detection and segmentation.Although traditional machine learning methods have shown superiority in the test results,this research aims to realize the full automation of image diagnosis,therefore selects the 3D-Res Net model to realize the tumor diagnosis.Due to the huge amount of parameters in deep learning model,there are certain requirements for the operating environment.In this research,the algorithm model is deployed on a remote workstation,the calculation is performed by a computer with powerful computing power,and pass diagnosis result through network communication.The solution not only can obtain the high-precision results predicted by the deep learning algorithm,but also helps the promotion of intelligent aided diagnosis software in practical applications by reducing the requirements for the operating environment.