Abstract
Purpose :
With little development of artificial intelligence(AI) for orbital CT scans and Graves’ disease, there is still a lack of AI exploration for dysthyroid optic neuropathy(DON). Therefore, we propose an AI model which assists in the autonomous diagnosis of DON by coronal orbital CT scans, using convolutional neural networks.
Methods :
A mobile convolutional neural networks model was constructed and trained on coronal CT scans of 50 patients with Graves’ disease. Of these, 25 CT scans were of patients without DON, and 25 CT scans were of patients with DON. Four to six consecutive coronal scans were retrieved between the posterior globe and orbital apex for each patient, yielding 116 images in Graves’ group, and 93 images in the DON group. A total of 209 serial coronal CT slices from each CT scans were utilized for training and inference of the AI model. The MobileNetV2 architecture, with pre-train by the BVLC model, formed the basis of the AI model which was trained to identify CT scans with or without DON.
Ground truth was determined by comprehensive clinical examination (visual acuity, pupillary exam, intraocular pressure, Hertel measurement, color vision, contrast sensitivity, visual field testing, CAS scores, optical coherence tomography, optical coherence tomography angiography, computed tomography (CT) scans, magnetic resonance imaging, and visual evoked potential/response, etc.) with a degree of confidence such that DON patients all underwent subsequent surgical orbital decompression.
Results :
The prediction accuracy is 95.215% in MobileNetV2, while AlexNet cannot reach convergence under a similar training process. Overfitting is suspected.
Conclusions :
Our preliminary results suggest that MobileNetV2 may assist in the detection of Dysthyroid Optic Neuropathy in Graves’ disease via autonomous analysis of coronal serial orbital CT scans. Further independent testing is merited for validation.
This abstract was presented at the 2023 ARVO Annual Meeting, held in New Orleans, LA, April 23-27, 2023.