基于深度学习的甲状腺结节自动识别方法在超声图像中的应用

Application of Automatic Thyroid Nodule Recognition Based on Deep Learning in Ultrasonic Image

目的构建和验证一个用于甲状腺结节自动识别的深度学习模型,旨在提高甲状腺结节识别和诊断水平。方法从超声数据库选取2013年1月至2018年1月期间6321张甲状腺图像,其中确诊为多发结节的2000张和确诊为单个结节的1200张用于深度学习模型训练,其他未确诊甲状腺图像3121张用于深度学习模型验证并提交给4名临床医师进行诊断,最后进行统计分析。结果深度学习方法在阳性预期率、阴性预期率、诊断敏感性、诊断效率和诊断特异性指标上都超过了超声医师。深度学习方法的阳性预期率比高年资超声医师高10.00%,阴性预期率高5.02%,诊断效率高10.24%。结论本研究构建的深度学习模型用于甲状腺结节的诊断具有较高的准确率,可在超声诊断甲状腺检查中辅助医师进行实时诊断。深度学习方法应用于超声影像的甲状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 thyroid nodules in order to improve the level of recognition and diagnosis of thyroid nodules. Methods A total of 6321 thyroid images from January 2013 to January 2018 were selected, of which 2000 images diagnosed as multiple nodules and 1200 images diagnosed as single nodules were used for deep learning model training, the other 3121 images were used for deep learning model validation and were submitted to 4 clinicians for diagnosis. Finally, statistical analysis was carried out. Results Deep learning method surpassed ultrasound physicians in positive expectation rate, negative expectation rate, diagnostic sensitivity, diagnostic efficiency and diagnostic specificity. The positive expectation rate of deep learning method was 10.00% points higher than that of senior ultrasound doctors, the negative expectation rate was 5.02% points higher, and the diagnostic efficiency was 10.24% points higher. Conclusion The deep learning model constructed in this study has high accuracy in the diagnosis of thyroid nodules, and can assist physicians in real-time diagnosis of thyroid nodules in ultrasound diagnosis. It is feasible to apply deep learning method to the clinical diagnosis of thyroid nodules in ultrasound images.