Diagnosing Retinal Vascular Occlusion using Artificial Intelligence-assisted Imaging Classification at High Accuracy

Abstract

Purpose : Retinal vascular occlusions (RVOs) are common causes of visual impairment. Timely diagnosis and acute management of RVOs are crucial to prevent irreversible vision loss and achieve optimal recovery. We explored the integration of cloud-based artificial intelligence (AI) as a complementary tool to assist ophthalmologists in swiftly identifying RVOs based on distinct patterns in fundus images.

Methods : We leveraged a curated dataset of 981 (500 normal, 481 with RVOs) publicly available retinal images sourced from Kaggle. The dataset included images representing various stages and types of RVOs. These were divided into a training set of 785 images, a validation set of 98 images, and a test set of 98 images. The automated machine learning (AutoML) process was utilized to streamline hyperparameter optimization and AI deployment. The model underwent training on Google Cloud over a duration of 2 hours and 28 minutes. Performance evaluation metrics encompassed accuracy, precision, recall (sensitivity), specificity, and the F1 score.

Results : The AI model exhibited robust performance, achieving an accuracy of 99%, precision of 98%, recall of 100%, specificity of 98%, and an F1 score of 99%. Furthermore, the model demonstrated an exceptional Area Under the Curve (AUC) value of 0.976, underscoring its discriminative prowess in identifying the presence of RVOs. Figure 1 shows the confusion matrix of the predictions made by the model on the test set. Example images from the test set and the confidence of the model’s predictions are shown in Figure 2. These findings underscore the model's potential clinical utility for accurate and prompt diagnosis of RVOs.

Conclusions : We successfully demonstrated the feasibility and efficiency of deploying a cloud-based AI platform to detect RVOs from retinal images. The high accuracy, AUC, precision, recall, specificity, and F1 score achieved with our model highlight its promise as a valuable tool to augment and expedite clinical decision making. This cloud-based AI model with AutoML features could be easily and widely applied, which will enable early diagnosis and treatment of RVOs in medically under-resourced areas. Therefore, it could significantly improve the prognosis for patients with RVOs and alleviate burden on the healthcare system.

This abstract was presented at the 2024 ARVO Annual Meeting, held in Seattle, WA, May 5-9, 2024.

 

Figure 1: Confusion Matrix Assessing Model Performance

 

Figure 2: Model Classification for RVO