Purpose

Thyroid-associated ophthalmopathy (TAO) might lead to blindness and orbital deformity. The early diagnosis and treatment are conducive to control disease progression, but currently, there is no effective screening method. The present study aimed to introduce an artificial intelligence (AI) model for screening and testing the model with TAO patients under clinical conditions.

Methods

A total of 1435 computed tomography (CT) scans were obtained from the hospital. These CT scans were preprocessed by resampling and extracting the region of interest. CT from 193 TAO patients and 715 healthy individuals were adopted for three-dimensional (3D)-ResNet model training, and 49 TAO patients and 178 healthy people were adopted for external verification. Data from 150 TAO patients and 150 healthy people were utilized for application tests under clinical conditions, including non-inferiority experiments and diagnostic tests, respectively.

Results

In the external verification of the model, the area under the receiver operating characteristic (ROC) curve (AUC) was 0.919, indicating a satisfactory classification effect. The accuracy, sensitivity, and specificity were 0.87, 088, and 0.85, respectively. In non-inferiority experiments: the accuracy was 85.67% in the AI group and 84.33% in the resident group. The model passed both non-inferiority experiments (p = 0.001) and diagnostic test (the AI group sensitivity = 0.87 and specificity = 0.84%).

Conclusions

A promising orbital CT-based TAO screening AI model was established and passed application tests under clinical conditions. This may provide a new TAO screening tool with further validation.

中文翻译：

---

甲状腺相关性眼病的人工智能CT筛选模型和临床条件下的测试

目的

甲状腺相关性眼病（TAO）可能导致失明和眼眶畸形。早期诊断和治疗有利于控制疾病的进展，但目前尚无有效的筛查方法。本研究旨在介绍一种人工智能（AI）模型，用于在临床条件下对TAO患者进行筛选和测试。

方法

从该医院总共进行了1435次计算机断层扫描（CT）扫描。这些CT扫描通过重新采样并提取感兴趣区域进行预处理。对来自193名TAO患者和715名健康个体的CT进行了三维（3D）-ResNet模型训练，并采用了49名TAO患者和178名健康人进行外部验证。来自150名TAO患者和150名健康人的数据分别用于临床条件下的应用测试，包括非自卑性实验和诊断测试。

结果

在模型的外部验证中，接收器工作特性（ROC）曲线（AUC）下的面积为0.919，表明令人满意的分类效果。准确性，敏感性和特异性分别为0.87、088和0.85。在非自卑性实验中：AI组的准确率为85.67％，常住组的准确率为84.33％。该模型通过了非劣效性实验（p  = 0.001）和诊断性测试（AI组敏感性= 0.87，特异性= 0.84％）。

结论

建立了有前途的基于轨道CT的TAO筛查AI模型，并在临床条件下通过了应用测试。这可以提供具有进一步验证的新的TAO筛选工具。