Abstract
Purpose :
Thyroid eye disease (TED) and orbital myositis (OM) are two distinct orbital pathologies which may present with overlapping clinical presentations. Differentiation of these diseases is crucial in order to initiate appropriate treatment and to prevent complications like permanent disability or vision loss. This study aims to develop a deep learning (DL) model to accurately classify TED and OM based on orbital computed tomography (CT) images.
Methods :
A retrospective review was conducted on patients who underwent dedicated orbital CT scans over a 12-year period at a single academic institution. Patients diagnosed with OM and TED by an oculoplastics provider were included, while those with other diagnoses (i.e. trauma, tumors, or other inflammatory processes), or with prior orbital surgeries (i.e. bony decompressions) were excluded. TED was categorized into mild TED and severe TED (defined as those with clinical signs of optic neuropathy). Patients with no orbital disease were normal controls. Orbital CTs were preprocessed and adopted for the VCG-16 network to develop and train a deep-learning (DL) model to distinguish patients with no disease, mild TED, severe TED, and orbital myositis. The DL model was trained on 80% of the data and tested on 20%. Additional subset analysis was performed to classify TED from myositis based on lateral rectus involvement.
Results :
A total of 1628 single coronal slices from 205 CT scans (31 controls, 83 mild TED, 40 severe TED, and 51 orbital myositis) were included. The DL model distinguished TED (mild and severe) from OM, with an accuracy rate of 98.4% (AUC 0.999). The model distinguished mild TED from OM with an accuracy of 98% (AUC 0.999), and severe TED from OM with an accuracy of 98% (AUC 0.996). Subset analysis of the lateral rectus (LR) only showed the model could distinguish TED from OM with an accuracy of 99.5% (AUC 0.994), and subset analysis of non-LR involving patients showed an accuracy of 95.6% (AUC 0.996).
Conclusions :
The DL model developed in this study demonstrates promising results in accurately classifying TED and orbital myositis. The continued development and validation of this DL model can aid in improving the efficiency and accuracy of diagnosing and triaging referrals for orbital conditions.
This abstract was presented at the 2024 ARVO Annual Meeting, held in Seattle, WA, May 5-9, 2024.