Transforming Ocular Health: A Cloud-Driven AI Approach to Drusen Detection in OCT Scans

Abstract

Purpose : Drusen, extracellular deposits between retinal pigment epithelium and Bruch's membrane, are vital indicators in ocular health, often linked to age-related macular degeneration (AMD), affecting 196 million people globally. This study aims to develop an artificial intelligence (AI) model differentiating Drusen and normal Optical Coherence Tomography (OCT) scans.

Methods : OCT scans served as the foundation for training the model, focusing on its ability to discern subtle features indicative of Drusen pathology. Training focused on discerning subtle Drusen features using Google's cloud-based services. The utilization of Google's platform underscored the advantages of cloud-based services for computationally intensive tasks, allowing for streamlined and resource-efficient model development. Utilizing Google's platform, the model underwent efficient 2-hour and 8-minute training, cost-free and carbon-neutral. Employing a dataset from Kaggle.com consisting of 5932 OCT scans where 2935 were OCT scans with Drusen and 2997 were OCT scans with normal pathology. The data was randomly split into training (80%), validation(10%), and testing sets (10%).

Results : The AI model successfully differentiates Drusen and normal OCT scans, demonstrating high accuracy and precision. The model was evaluated using the following metrics: accuracy (95%), precision (97.98%), recall (97%), specificity (98%), F1-score (97.4%), and AUC (0.996). The high AUC emphasizes the model’s efficacy in capturing nuanced Drusen patterns, providing a comprehensive evaluation of its discriminative capabilities.

Conclusions : This investigation successfully showcases AI model development for Drusen detection in OCT scans, emphasizing the advantages of cloud-based resources for efficient and cost-effective training. Integrating a publicly available dataset with careful data partitioning enhances accessibility. The remarkable AUC and other metrics suggest the model's potential as a valuable tool in identifying early Drusen-related conditions amidst the global health burden posed by AMD.

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

 

Figure 1: Confusion Matrix Assessing and Classifying Classes of the Model

 

Figure 2: Classification of Drusen Pathology from the Model