Research On Assisted Diagnosis Of Microcalcification Clusters Based On Cascade RCNN And Digital Breast Tomosynthesis

Nowadays,breast cancer has the highest incidence rate among female malignant tumors.Research shows that if patients can screen out breast cancer at an early stage,timely treatment can greatly reduce the threat to women’s life and health.As one of the common signs of breast cancer,it is very necessary to detect microcalcifications to assist doctors in diagnosis and treatment of breast cancer.However,the common two-dimensional X-ray imaging in breast examination has many shortcomings,such as tissue overlap and structural noise.Digital Breast Tomography(DBT)is a three-dimensional imaging method that can provide image information across various sections of the breast,effectively avoiding the above problems.In terms of auxiliary diagnosis and treatment for DBT images,traditional machine learning methods have some shortcomings,such as cumbersome steps,manual extraction of microcalcification cluster features,manual labeling of regions of interest,establishment of classifier implementations,low detection efficiency,and poor generalization.In order to solve the above problems,this study aims at the development of an auxiliary diagnosis system,using deep learning technology to achieve auxiliary diagnosis of microcalcification clusters in DBT images.The main work contents are as follows:1.This study is based on an improved Cascade RCNN model to achieve the detection of microcalcifications.The SE-Res Ne Xt-50 obtained by fusing the Squeeze-and-Excitation module with the residual module Res Ne Xt-50 is used as the backbone network,replacing the original Res Net-50,increasing the receptive field of each network layer,providing stronger feature expression capabilities for the generation of RPN network candidate regions in the subsequent Cascade RCNN.2.Each module in SE-Res Ne Xt-50 has completed feature extraction for microcalcification cluster lesions,but this method has some shortcomings.The fundamental reason is that regular rectangles are used to extract features,which can lead to incomplete collection of some relatively complete features and loss of many features.To solve this problem,this study selects deformable convolution DCNv2 to preserve more complete features.3.Traditional FPN integrates deep output feature maps into shallow networks without substantially increasing the computational complexity of the original model.Through upsampling of deep features and top-down connection of shallow features,shallow networks retain target location information in the image and provide more semantic information.Because most microcalcification cluster targets are relatively small and shallow networks contain more feature information,in order to increase the ability to extract features from microcalcification clusters,this study improved FPN by fusing shallow feature maps into deep layers.Non maximum suppression of NMS can easily lead to missed detection.This study uses Soft-NMS instead of NMS to solve this problem.4.In order to solve the problem of insufficient DBT data sets in Jiangxi Third Class Hospital,this study first conducted data enhancement processing for the data sets,and then adopted a model based migration approach in migration learning.First,pre training was conducted on two-dimensional molybdenum target images in CBIS-DDSM data sets,and then the pre-trained models were migrated to Jiangxi Third Class Hospital DBT data sets for fine tuning training,effectively improving the detection effect.