Deep learning models for predicting clinical activity and characteristics in thyroid eye disease using magnetic resonance imaging.
Authors
Affiliations (2)
Affiliations (2)
- Department of Ophthalmology and Visual Sciences, Yale School of Medicine, New Haven, Connecticut, USA.
- Department of Ophthalmology, Private Practice, Beverly Hills, California, USA.
Abstract
This project aims to develop and evaluate deep learning models using orbital magnetic resonance imaging for the prediction of continuous clinical activity score and key patient characteristics in thyroid eye disease. The publicly available TOM500 dataset, consisting of orbital magnetic resonance imaging scans and clinical data from 500 thyroid eye disease patients, was split into training (<i>n</i> = 360), validation (<i>n</i> = 100), and test (<i>n</i> = 40) sets. A ResNet-50 convolutional neural network pretrained on ImageNet was adapted for four tasks: clinical activity score, age, smoking status, and sex prediction. Each slice was converted into a 3-channel input: the raw scan, the corresponding segmentation mask, and a weighted average. Model performance was evaluated using mean absolute error and accuracy. The clinical activity score prediction model achieved a mean absolute error of 1.05. The age model achieved a mean absolute error of 6.90 years, while the sex and smoking status models achieved an accuracy of 93% and 80%, respectively. The area under the receiver operating characteristic for the sex and smoking status models was 0.974 and 0.747, respectively. This study highlights the use of artificial intelligence in thyroid eye disease. It offers a granular assessment of disease progression while demonstrating the feasibility of extracting clinical characteristics directly from orbital magnetic resonance imaging.