Study on the diagnostic efficacy of multimodal imaging based on deep learning in breast ductal carcinoma in situ

WANG Juanjuan; MA Jie; YI Chunyan; ZHAN Meimao; ZHANG Shixin; CHENG Guanxun

Abstract：Objective To investigate the diagnostic efficiency of difference from full field digital mammography(FFDM),magnetic resonance imaging(MRI)and ultrasound(US)in breast ductal carcinoma in situ(DCIS),and the role of artificial in-telligence system in assisting junior radiologists to read films.Methods The cases of DCIS confirmed by surgery and pathology were collected and analyzed retrospectively.91 female patients with breast DCIS were examined by FFDM,US and MRI before operation.91 patients underwent FFDM examination,and the diagnosis was assisted by senior radiologists and junior radiologists combined with artificial intelligence system.The 79 patients underwent US examination and 49 patients underwent MRI examina-tion,of whom the 42 patients combined FFDM,US and MRI.All lesions were classified by the ACR 5th Edition Breast Imaging Reporting and Data System(BI-RADS)in 2013.We calculated their sensitivity and false negative rate,and evaluated their diag-nostic efficacy by chi square test;The related factors affecting the sensitivity of three image examination methods were analyzed.The 91 cases of FFDM images were detected by junior radiologists with the aid of artificial intelligence(Mammo AI),and their sensitivity and false negative rate were calculated.Chi square test was used to compare the diagnostic accuracy with the radiology senior medical doctors.Results The diagnostic sensitivityof FFDM,US and MRI was 78.0%(71/91),78.5%(62/79)and 95.9%(47/49),respectively.The difference was statistically significant(χ2=8.132,P<0.05).When the three methods were used together,the positive rate of DCIS was 97.6%(41/42).The sensitivity of 78.0%(71/91)by senior radiologists andprimary ra-diologistscombined with AI 75.8%(69/91)was not statistically significant(χ2=0.124,P>0.05).FFDM is easy to be misdiagnosed due to the influence of dense glands and non-calcified lesions,while non mass and calcified lesions were the main reasons for our misdiagnosis;The 2 cases of MRI underestimation were mainly affected by the enhancement of lesions and limited diffusion.Conclusions The sensitivity of MRI in the diagnosis of DCIS is higher than that of FFDM and US,with lower false negatives and less missed diagnosis rate.The combined application of the three imaging examination methods can improve the diagnostic accuracy of DCIS and avoid missed and misdiagnosis.Mammo AI reduces radiologist viewing time and improves diagnostic effi-ciency of mammography images.