基于深度学习模型在乳腺结节钼靶中的应用

Application of breastnodules in mammographybased on deep learning

目的构建和验证一个用于乳腺结节自动识别的深度学习模型,旨在提高乳腺结节识别和诊断水平,在钼靶检查中辅助放射医师进行实时诊断。方法从数据库系统选取2013年1月至2019年1月2 813张乳腺钼靶图像,其中包括确诊单个结节1 200张,多发结节490张,用于深度学习模型训练。其他未确诊乳腺钼靶图像1 123张用于深度学习模型验证,同时提交给4名放射医师进行诊断,最后进行统计。结果根据测试对比结果,分别计算阳性预测值、阴性预测值、诊断敏感度、诊断效率和诊断特异度指标,机器学习方法在所有指标比较中都达到高年资医师水平。其中诊断敏感度比低年资放射医师高出10.39%,诊断特异度高出8.04%,诊断效率高出10.15%。结论本研究构建的深度学习模型用于乳腺结节的诊断水平与高年资放射医师相仿,可在乳腺钼靶检查中辅助放射医师进行实时诊断。机器学习方法应用于钼靶的乳腺结节的临床辅助诊断是可行的。

Objective To construct and validate a deep learning model for automatic recognition of breast nodules in order to improve the level of recognition and diagnosis of breast nodules,helping radiologist with real-time diagnosis in mammography. Methods Two thousand eight hundred and thirteen mammography images from January2013 to January 2019 were selected from the database system, including 1 200 single nodules and 490 multiple nodules for deep learning model training. A total of 1 123 undiagnosed mammography images were used for deep learning model validation, submitted to 4 radiologists for diagnosis, and finally counted. Results According to the test results, the positive predictive value, the negative predictive value, the diagnostic sensitivity, the diagnostic efficiency and the diagnostic specificity were calculated respectively. The machine learning method achieved the level of seniorradiologists in all the indexes. The diagnostic sensitivity was 10.39% higher than that of junior radiologists, the diagnostic specificity was 8.04% higher and the diagnostic efficiency was 10.15% higher. Conclusion The in-depth learning model constructed in this study is similar to that of senior radiologists in the diagnosis of breast nodules. and can assist physicians in real-time diagnosis of breast nodules by mammography. It is feasible to apply machine learning method to the clinical diagnosis of breast nodules inmammography images.