Intelligent Assistant Diagnosis Of Breast Architectural Distortion Based On DBT Images

Objectives:This study aims to explore the clinical application value of intelligence-assisted methods in the diagnosis of breast structural distortion lesions based on digital breast tomosynthesis images.The main objectives of the study were as follows:1.Comparison of Digital Breast Tomosynthesis(DBT)and Full-field digital mammography(DBT)FFDM)to detect Architectural distortion(AD),and the factors affecting the detection of AD lesions were discussed based on Logistics regression analysis.(2)A deep learning-based AD detection method is proposed.In DBT,the convolutional neural network is used to detect AD lesions,and the adaptive receptive field feature extraction method is used to improve the detection ability of atypical AD lesions,in order to solve the difficulty of detecting structural distortion lesions in DBT images.In DBT,the method of convolutional neural network is used to classify benign and malignant AD lesions,typical and atypical AD lesions,and the output results are used to evaluate the clinical effectiveness of the deep learning classification model.Materials and Methods:Part I:The clinical and imaging data of breast AD patients who met the inclusion and exclusion criteria in Nanfang Hospital of Southern Medical University from March 2014 to July 2022 were retrospectively collected,and a total of 1260 patients were included.The description and terms of AD in the FFDM/DBT extraction report were analyzed.According to the doctor’s description,the results were divided into FFDM and DBT detection,and only DBT detection,and the AD detection rate of the two methods was calculated.Then,all cases were classified as"typical" and "atypical" by two radiologists.To explore the differences in the detection and diagnosis of AD between FFDM and DBT with different clinical characteristics and different AD classifications.To explore the differences in the detection of different types of ADs by different seniority.The detection rate and consistency were calculated.Logistic regression analysis was used to evaluate the characteristics affecting the detection and diagnosis of AD.Part Ⅱ:Based on the data collected in Part I,307 AD cases were randomly selected by serial number random method.The age ranged from 20 to 74 years(mean ± standard deviation:45.71±7.78).There were 207 positive cases(with AD)and 100 negative cases(without AD).The data were divided into training set and independent test set.The true location of the AD on each side was marked by a radiologist using the location of the lesion recorded surgically as the gold standard.Based on the labeled DBT images,Gabor filtering and convergence degree were preprocessed,and all preprocessing results were merged as input.Then,the upper-and lower-layer information of AD lesion DBT images and anatomical prior knowledge were used as additional information to help the model reduce the number of false positive lesions,and a detection model based on deformable convolutional network was established,which was evaluated on an independent test set.The Free Receiver Operating Characteristic Curve(FROC)was used to determine the detection performance of the model.Part III:Based on the first part of the data set,404 AD cases were randomly selected according to the serial number method.Of these,294 were benign(positive)and 110 were malignant.136 cases were typical and 268 cases were atypical.In each DBT sequence,the location of the AD was labeled by a radiologist based on the location recorded by surgical pathology,which was used as the gold standard for model classification.According to the ratio of 7:3,the patients were randomly divided into training set and validation set.The VGG-16,Resnet-34 and Resnet-101 convolutional network structures were used to construct the classification models of benign and malignant,typical and atypical AD lesions on DBT images,and the classification models were evaluated on an independent test set.The VGG-16 network was selected as the baseline model,and the classification performance was determined by response Receiver Operating Characteristic Curve(ROC).Results:Part I:The detection rate of FFDM was 74.8%,and that of DBT was 99.5%.DBT had a better ability to detect AD,There were significant differences in the ability to detect AD among different seniority physicians,and the higher the seniority,the higher the detection rate.AD lesions showing"radiate at one point" were more likely to be diagnosed as typical AD(OR=684.72,95%CI:70.94-6608.51).AD lesions located in the "upper outer quadrant" are more likely to be diagnosed as atypical AD.Part II:The sensitivity of the baseline method(convergence-based method)for AD detection did not reach 80%.The Mean true positive fraction(MTPF)of the proposed optimization model was 0.664.The method based on context information and anatomical prior knowledge to reduce the number of False Positive(FP)is effective.In the optimization experiment of atypical AD lesion detection,the proposed optimization model has improved detection ability compared with the conventional convolution model.Part III:The three networks(VGG-16,Resnet-34,and Resnet-101)were validated in an independent test set.By comparing the classification performance of the three networks,Resnet-34 network had the best performance in the typical atypical classification experiment,with an ACC of 0.704 and an AUC of 0.615,and Resnet-34 was selected as the final classification network.In the classification experiment of benign and malignant AD,Resnet-101 network had the best performance with AUC of 0.844 and ACC of 0.785,and Resnet-101 was selected as the final classification network.Conclusion:(1)The ability to detect AD lesions of the two methods was higher than that of DBT,especially for atypical AD lesions.The ability to detect AD tended to increase with the increase of seniority.2."Focal shrinkage" and "increased central density" were unfavorable factors for AD detection,while "central region","middle 1/3 region" and "seniority-intermediate level" were favorable factors for AD detection.The AD detection model based on deep learning method can improve the detection rate of AD lesions in DBT images.The number of false positive lesions can be reduced by using the information of up-down direction and anatomical prior knowledge.(4)The adaptive receptive field method can effectively improve the detection ability of atypical AD lesions,indicating that global features can better describe the characteristics of atypical AD lesions.The classification model based on Resnet-34 and Resnet-101 network can be used as the AD lesion classification network.