基于眼肌语义分割和特征提取的甲状腺相关眼病诊断算法研究

Diagnosis of thyroid-associated ophthalmopathy with semantic segmentation and feature extraction

目的甲状腺相关眼病(thyroid-associated ophthalmopathy,TAO)是常见的眼病之一,通过CT图像进行诊断和筛查对治疗有着重要意义,但传统方法依赖有经验的医生对CT进行分析和诊断,尚无有效的自动化方法。为此本文提出一种可以从CT图像中自动提取特征进行TAO诊断的方法,辅助医生进行诊断。方法设计了Unet-Orbit分割网络对CT中的眼肌进行图像分割,随后采用影像组学工具(PyRadiomics)从分割结果中的眼肌区域提取数值化特征。为了更好地利用影像组学的特征,设计了一个特征提取网络,采用自动编码器框架。将不同的眼肌提取到的特征,通过特征合并和变换进一步得到一组新特征。最后采用来自上海交通大学医学院附属第九人民医院的1912个CT图像数据集,对使用原始影像组学特征的分类器与使用特征提取网络后的特征的分类器进行了比较。结果在医院数据集上,该模型的诊断准确率、灵敏度和特异性分别为87.34%、84.73%和89.96%。结论语义分割网络可以高效分割眼肌区域,特征提取网络得到的新特征可以提升多种不同分类器在TAO诊断的准确率,可能为TAO的诊断提供一个新工具。

Objective Thyroid-associated ophthalmopathy(TAO)is one of the common orbital diseases.The diagnosis and screening for TAO based on orbital CT are important for treatment.But currently,traditional methods rely on experienced doctors to analyze and diagnose CT images.There is no effective automatic screening method.In this study,we propose a method that can automatically extract features from CT images for TAO diagnosis to assist doctors in diagnosis.Methods Unet-Orbit model was designed segmenting the eye muscles in CT images.Then,a radiomics tool(PyRadiomics)was used to extract digital features from the eye muscle regions in the segmentation results.To effectively exploit the features from CT images,a feature extraction network was designed by using an autoencoder framework.The features extracted from different eye muscles were further combined and transformed to obtain a set of new features.Finally,the dataset used 1912 CT images,which were provided by Shanghai Ninth People’s Hospital,Shanghai Jiao Tong University School of Medicine.The study compared the classifier using raw radiomic features with the classifier using the new features after feature extraction network.Results On the hospital’s dataset,the accuracy,sensitivity,and specificity of the model were 87.34%,84.73%,and 89.96%.Conclusions The semantic segmentation networks can efficiently segment eye muscles,and the feature extraction network can be used for different classifiers to promote the diagnosis accuracy.This may provide a new TAO diagnosis tool.