Abstract
Purpose :
The early and accurate diagnosis of acute retinal necrosis syndrome (ARN) is crucial for effective clinical management and minimizing the risk of irreversible vision loss. This multicenter study developed a deep learning model for detection of uveitis and ARN.
Methods :
We developed and tested Swin Transformer-based deep learning model (DeepDrARN), using 5124 ultra-widefield color fundus photographs (UWFCFPs) of 908 subjects. Firstly, we trained and internally validated DeepDrARN on cohorts form Eye Hospital of Wenzhou Medical University, and then externally tested DeepDrARN on cohorts from Ningbo Eye Hospital. We further compared the performance of DeepDrARN with seven ophthalmologists in another independent cohort.
Results :
DeepDrARN showed robust diagnostic performance with sensitivities of 99.5% and 98.6%, specificities of 87.1% and 65.9%, areas under the receiver operating characteristic curves (AUC) of 0.996 and 0.973 for detecting uveitis, and sensitivities of 84.4% and 73.7%, specificities of 95.9% and 98.2%, AUC of 0.960 and 0.971 for detecting ARN in the internal and external validation cohorts, respectively. In another independent validation cohort, DeepDrARN achieved a superior diagnostic performance comparable with human ophthalmologists, with an average accuracy increase of 6.57% and 11.14% in detecting uveitis and ARN. The cost-benefit analysis demonstrated that DeepDrARN substantially reduces economic and time costs compared to the 'gold standard' for diagnosis, and increases efficiency and time benefits compared to conventional empirical diagnosis.
Conclusions :
Our study demonstrates the potential and superiority of deep-learning algorithms as convenient, fast and low-cost tools in the early detection and accurate diagnosis of ARN.
This abstract was presented at the 2024 ARVO Annual Meeting, held in Seattle, WA, May 5-9, 2024.