摘要目的 研发基于深度神经网络的智能识别甲状腺功能状态的诊断模型.方法 选取2016年5月至2018年6月1616例(男283例,女1333例,平均年龄52岁)临床已确诊受检者的甲状腺平面显像图,其中甲状腺正常图像299例,甲状腺功能亢进症(简称甲亢)图像876例,甲状腺功能减退症(简称甲减)图像441例.利用2种深度神经网络模型AlexNet和深度卷积生成对抗网络(DCGAN),应用深度学习算法对1000例训练集样本进行特征提取和训练,对616例测试集样本进行效能验证,对2种模型的验证结果分别进行分析,采用Kappa检验进行一致性分析,并分析智能诊断模型的时间优越性.结果 AlexNet模型平均诊断时间为1 s/例,其对甲状腺功能正常、甲亢和甲减的分类判别准确性分别为82.29％(79/96)、94.62％(369/390)、100％(130/130),分类结果与确诊结果的一致性检验Kappa值为0.886(P<0.05);DCGAN模型平均诊断时间为1 s/例,其对甲状腺功能正常、甲亢和甲减的分类判别准确性分别为85.42％(82/96)、95.64％(373/390)、99.23％(129/130),分类结果与确诊结果的一致性检验Kappa值为0.904(P<0.05).结论 深度神经网络智能诊断模型可快速判别甲状腺平面显像中甲状腺的功能状态,识别准确性较高,为甲状腺平面显像图审阅提供了新方式.

abstractsObjective To develop a diagnostic model based on deep neural network for intelligent discrimination of thyroid function. Methods A total of 1616 patients ( 283 males, 1333 females, average age:52 years) who underwent thyroid imaging between May 2016 and June 2018 were selected. According to the clinical diagnosis, the 1616 cases included 299 normal thyroid cases, 876 hyperthyroidism cases and 441 hypothyroidism cases. Feature extraction and learning training were performed on 1000 training set sam-ples by two deep neural network models ( AlexNet;deep convolution generative adversarial networks ( DCGAN) ) using deep learning algorithm. Performance verifications were implemented on 616 test set samples. The con-sistency between the verification results of the two models and the clinical diagnosis was analyzed by Kappa test. Meanwhile, the time advantage of the intelligent diagnosis models was analyzed. Results The average diagnostic time of AlexNet model was 1 s/case, and the classification accuracy for normal thyroid, hyperthy-roidism, hypothyroidism were 82.29％(79/96), 94.62％(369/390), 100％(130/130), respectively. The Kappa value between results of AlexNet model and clinical diagnosis was 0.886 ( P<0.05) . The average di-agnostic time of DCGAN model was 1 s/case, and the classification accuracy for normal thyroid, hyperthy-roidism, hypothyroidism were 85.42％(82/96), 95.64％(373/390), 99.23％(129/130), respectively. The Kappa value between results of DCGAN model and clinical diagnosis was 0.904 ( P<0.05) . Conclusion The deep neural network intelligent diagnosis model can quickly determine the functional status of thyroid gland in thyroid imaging, and it has a high recognition accuracy, thus providing a new method for thyroid image review.