Abstract
Purpose :
Thyroid eye disease (TED) is a rare autoimmune disease that leads to enlargement of the extraocular muscles, fat and connective tissue volume. Given the characteristic orbital involvement, computed tomography (CT) has been widely adopted to aid in the diagnosis and monitoring of TED. Our present study aims to enhance the neural network-based method in screening for and assessing the severity of TED.
Methods :
Patients were divided into three subgroups based on clinical diagnoses and radiographic findings. In the control group, patients were seen at the Massachusetts Eye and Ear Oculoplastics clinics and underwent an orbit CT scan for a presumed orbital process but were not found to have any orbital pathology. In group two (Mild TED), patients were diagnosed with thyroid eye disease but no evidence of compressive optic neuropathy. In group three (Severe TED), patients were diagnosed with thyroid eye disease plus features of compressive optic neuropathy. The raw dataset included a total of 885 cross-section 2D images from the seventy-two CT scans in coronal view, which was resampled and adapted to focus on the eye as the region of interest. There were 20 eyes in the control group, 60 eyes in the Mild TED group, 64 eyes in the severe TED group.
The visual geometry group from Oxford (VGG16) model is a convolutional neural network pre-trained on ImageNet dataset, a collection with over 14 million images in 22,000 categories. All the TED datasets were trained using the VGG16 model 100 epochs. There were 628 images used in the training dataset: 231 images from the control group, 200 from group 2, and 197 from group 3. 157 images were used for the testing dataset: 50 images from group 1, 52 from group 2, and 55 from group 3.
Results :
The overall prediction accuracy is 94.27%. Two images from the control group were misclassified as group 2 (Mild TED), and six images from Mild TED were misclassified as the normal control group. Images from group 3, the severe TED group with compressive optic neuropathy was never misclassified as normal or Mild TED.
Conclusions :
Neural network-based analytic AI models can not only help diagnose TED and screen for disease severity for TED entirely based on CT scans, which may allow for automated screening of TED and timely referral of patients with compressive optic neuropathy from TED severe.
This abstract was presented at the 2023 ARVO Annual Meeting, held in New Orleans, LA, April 23-27, 2023.