基于深度学习的乳腺X线摄影钙化检出系统评估

Evaluation of mammography calcification detection system based on deep learning

探讨基于深度学习的乳腺X线影像钙化检出系统的价值.方法回顾性分析2013年1月至12月解放军总医院第五医学中心南院区乳腺X线常规检查1431例患者的5488幅影像,每例检查均拍摄头尾位(CC)及内外斜位(MLO)图像.通过低年资医师A独立阅片、高年资医师B审核的方式,建立钙化检出的参考标准.采用χ^2检验研究不同因素(钙化形态、钙化分布、分类、美国放射学院腺体构成分类、患者年龄)对于深度学习和医师A的影响.结果深度学习对所有钙化的敏感性96.76%(7649/7905),假阳性平均每幅影像1.04(5706/5488)个,平均每例检查3.99(5706/1431)个,假阳性率为42.73%(5706/13355).深度学习和医师A对于典型良性和可疑恶性间钙化、不同形态钙化的诊断差异均有统计学意义(P均<0.05).深度学习对于不同分布钙化、BI-RADS分类、美国放射学院腺体构成钙化的敏感性差异无统计学意义(P>0.05),而医师A的差异有统计学意义(P<0.05).深度学习和医师A对于不同年龄下钙化检出的敏感性差异均无统计学意义(P>0.05).结论基于深度学习的乳腺X线影像钙化检出系统具有很高的敏感性以及一定的稳定性,可以有效减少阅片流程中钙化,尤其是可疑恶性钙化的漏检.

Objective To evaluate the performance of a deep learning(DL)based mammogram calcification detection system.Methods Screening digital mammographic examinations with standard cranio-caudal(CC)and medio-lateral oblique(MLO)views were performed in 1431 women(5488 mammogram images)who were enrolled between January and December in 2013.The DL system and a radiologist detect calcifications separately,and then both results are reviewed by a moreexperiencedradiologist.Sensitivities of the DL model and radiologist were compared.Different calcification morphology,distribution,BI-RADS categories,breast density and patient age were investigated by χ^2 tests.Results For DL system,sensitivity of all kinds of calcifications were 96.76%(7649/7905).The average false positive was 1.04 per image(5706/5488),3.99 per case(5706/1431).The false positive rate was 42.73%(5706/13355).There was no significant differences for DL system with different calcification distribution,BI-RADS categories,breast densities and patient ages(P>0.05).Conclusion Deep learning based mammogram calcification detection system shows high sensitivity and stability,which may help to reduce the missing rate of calcification(especially the suspicious ones).