Deep Learning Models for Predicting Clinical Characteristics in Thyroid Eye Disease Using Orbital MRI Images
Author: Sriya Rajyam
Base Hospital / Institution: Yale School of Medicine
Abstract ID: 25-318
Purpose
There is high inter-observer variability in assessing TED on imaging, highlighting the need for objective assessment. Prior AI studies have used orbital MRI to predict binary disease activity, rather than a continuous clinical activity score (CAS). The aim is to develop deep learning models capable of predicting CAS, age and smoking status in TED patients using a comprehensive orbital MRI dataset.
Methods
The publicly available TOM500 dataset by Zhang et al., consisting of orbital MRI scans and clinical data from 500 TED patients, was split into training (n=360), validation (n=100), and test (n=40) sets. A ResNet-50 convolutional neural network pretrained on ImageNet was adapted for three tasks: CAS regression, age regression and smoking status classification.
Each MRI slice was converted into a 3-channel input: the raw MRI image, the corresponding segmentation mask, and a weighted MRI (raw MRI × mask), enabling the model to learn from both global anatomy and localized disease regions. To improve performance, weighted training samples using inverse class frequency were employed to address the underrepresentation of high CAS and reduce overfitting. Model performance was evaluated using mean absolute error (MAE) for CAS and age, and accuracy for smoking status.
Results
The CAS prediction model achieved a mean absolute error (MAE) of 1.05, suggesting the model can estimate disease activity within approximately one point on the CAS scale. The age model achieved a MAE of 6.90 years and the smoking status model achieved an accuracy of 75%.
Conclusion
This study is the first to use TOM500 to predict CAS, age and smoking status. It offers potential for more objective monitoring of disease progression and reduction of interobserver variability, while demonstrating the feasibility of extracting clinical characteristics from MRI images. It is further novel in its use of a relatively large dataset, segmentation-informed inputs, and multi-variable prediction. The study serves as a proof of concept for AI’s potential in TED, with next steps involving further exploration of radiographic biomarkers and model optimization.
Additional Authors
| First name | Last name | Base Hospital / Institution |
|---|---|---|
| Yonca | Arat | Yale School of Medicine |
| Shoaib | Ugradar | Department of Orbit and Oculoplastic Surgery Private Practice |
| Michelle | Maeng | Yale School of Medicine |
