基于深度学习多模态影像学检查对乳腺导管原位癌的诊断效能分析

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

目的探讨全视野数字乳腺X线摄影(FFDM)、超声(US)及磁共振(MRI)检查在乳腺导管原位癌(DCIS)中的诊断效能,以及乳腺癌智能辅助诊疗系统辅助放射科初级医师阅片的作用。方法选取经手术病理证实为DCIS的患者91例。91例患者均接受FFDM检查,且诊断时分别由放射科高级医师和放射科初级医师结合人工智能系统辅助诊断阅片评估,79例患者接受US检查,49例患者接受MRI检查,其中42例患者联合了FFDM、US及MRI三种检查。所有病灶均采用2013年ACR第五版乳腺影像报告和数据系统(BI-RADS)进行分类,计算其灵敏度和假阴性率,采用卡方检验,评价其各自的诊断效能;分析3种影像学检查方法灵敏度的相关影响因素。91例FFDM图像由2位放射科初级医师借助《乳腺癌智能辅助诊疗系统》(简称:Mammo AI)对病灶进行检出,并对其进行良恶性的评估,亦计算其灵敏度及假阴性率,采用卡方检验,与放射科高级医师独立阅片进行比较,评价其诊断效能。结果FFDM、US及MRI的诊断灵敏度分别为78.0%(71/91)、78.5%(62/79)和95.9%(47/49),差异有统计学意义(χ^(2)=8.132,P<0.05),三种影像学检查方法联合使用时,DCIS的诊断阳性率97.6%(41/42)。放射科高级医师和Mammo AI辅助放射科初级医师对FFDM的诊断灵敏度分别为78.0%(71/91)和75.8%(69/91),差异无统计学意义(χ^(2)=0.124,P>0.05)。FFDM易受致密型腺体及非钙化型病变影响而漏误诊,而非肿块型和钙化型病变是US检查漏误诊的主要原因;MRI检查低估的2例主要受病灶强化情况及弥散受限情况的影响。结论MRI诊断DCIS的灵敏度高于FFDM和US,假阴性率低,漏诊率少,三种影像检查方法联合应用可提高DCIS的诊断准确率,避免漏误诊。Mammo AI降低了放射科医师的阅片时间,提高对乳腺X线摄影检查的诊断效率。

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 intelligence 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 examination,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 diagnostic 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(χ^(2)=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(χ^(2)=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.