基于眼球特征提取的甲状腺眼病预测方法研究
Research on Prediction Method of Thyroid-Associated Ophthalmopathy Based on Eyeball Feature Extraction

目的：眼球突出度的计算对甲状腺相关性眼病(Thyroid-Associated Ophthalmopathy, TAO)诊断非常有价值，但受限于传统眼球突出度测量工具的缺陷和眼科医生主观上的诊断差异，从CT影像中提取的多种特征对TAO进行预测可以避免上述误差。方法：使用深度学习方法建立眼球分割模型，从CT影像中分割出眼球，设计算法进行自动化测量眼球突出度，对眼球分割结果进行影像组学特征提取，利用不同特征组合对TAO进行预测，并观察比较结果差异。结果：眼球分割模型在测试集上的分割Dice系数为0.935，使用自动算法测量的眼球突出度与提取的眼球影像组学特征结合共同预测TAO的准确率优于单独使用眼球突出度或单独使用眼球组学特征对TAO进行预测的准确率。结论：自动化眼球突出度测量与影像组学特征结合高于单项指标预测效能，基于眼球影像特征提取对预测TAO的结果证明CT影像上不能被眼科医生直接发现的影像组学特征信息对预测结果有正面作用。
Objective: The calculation of proptosis is very valuable for the diagnosis of Thyroid-Associated Ophthalmopathy (TAO), but limited by the insufficiency of traditional tools for measuring exophthalmos and the subjective diagnostic differences of ophthalmologists. The various features extracted from CT images in this paper are useful for TAO to make predictions. Methods: A deep learning method was used to establish an eyeball segmentation model, and the eyeballs were segmented from CT images. An algorithm was designed to automatically measure the exophthalmos, and extract the features of eyeball segmentation results. Different feature combinations were used to predict TAO, and then observed and compared the difference in results. Results: The segmentation dice of the eye segmentation model on the test set was 0.935. The combination of the exophthalmos measured by the automatic algorithm and the extracted radiomic features together predicted the accuracy of TAO better than the exophthalmos or the radiomic features alone. Conclusion: The combination of automated proptosis measurement and radiomic features is higher than the prediction performance of a single index, and the prediction of TAO based on feature extraction proves that the radiomic features information on CT images that cannot be directly discovered by ophthalmologists has a positive effect on the prediction results.