AI system is helpful for diagnosis of CLE (table1&2)

Table I. Our development set consists of 3007 patient cases with 924 clinical images labeled by dermatologists, of which 197 patients were included in our reader study. Table II. Evaluation metrics, including accuracy, specificity, sensitivity and the kappa coefficient, were utilized to evaluate the diagnostic performance of our model. Experimental results showed that our model achieved the best performance when compared with three widely used CNN models, including SE-ResNeXt101-32x4d, SE-ResNet101(3) and Inception-v3 (4). The AIDDA outperformed these existing models, obtaining the highest AUC (0.973). The AUCs of SE-ResNeXt101-32x4d, SE-ResNet101, and Inception-v3 were 0.968, 0.965, and 0.964, respectively.

表1 本研究的开发集包含3007例患者病例，附带924张经皮肤科医师标注的临床影像，其中197例患者被纳入本研究的阅片者试验。
表2 本研究采用准确率、特异度、灵敏度及科恩Kappa系数作为评估指标，用以评价所构建模型的诊断性能。实验结果表明，与SE-ResNeXt101-32x4d、SE-ResNet101(3)及Inception-v3(4)这三款广泛应用的卷积神经网络（Convolutional Neural Network, CNN）模型相比，本研究的AIDDA模型取得了最优性能，其受试者工作特征曲线下面积（Area Under Curve, AUC）值为0.973，为所有对比模型中的最高值。SE-ResNeXt101-32x4d、SE-ResNet101与Inception-v3的AUC值分别为0.968、0.965及0.964。