Optimizing Ocular Pathology Classification with CNNs and OCT Imaging: A Systematic and Performance Review

Abstract

Vision loss due to chronic-degenerative diseases is a primary cause of blindness worldwide. Deep learning architectures utilizing optical coherence tomography images have proven effective for the early diagnosis of ocular pathologies. Nevertheless, most studies have emphasized the best outcomes using optimal hyperparameter combinations and extensive data availability. This focus has eclipsed the exploration of how model learning capacity varies with different data volumes. The current study evaluates the learning capabilities of efficient deep-learning classification models across various data amounts, aiming to determine the necessary data portion for effective clinical trial classifications of ocular pathologies. A comprehensive review was conducted, which included 295 papers that employed OCT images to classify one or more of the following retinal pathologies: Drusen, Diabetic Macular Edema, and Choroidal Neovascularization. Performance metrics and dataset details were extracted from these studies. Four Convolutional Neural Networks were selected and trained using three strategies: initializing with random weights, fine-tuning, and retraining only the classification layers. The resultant performance was compared based on training size and strategy to identify the optimal combination of model size, dataset size, and training approach. The findings revealed that, among the models trained with various strategies and data volumes, three achieved 99.9% accuracy, precision, recall, and F1 score. Two of these models were fine-tuned, and one used random weight initialization. Remarkably, two models reached 99% accuracy using only 10% of the original training dataset. Additionally, a model that was less than 10% the size of the others achieved 98.7% accuracy and an F1 score on the test set while requiring 100 times less computing time. This study is the first to assess the impact of training data size and model complexity on performance metrics across three scenarios: random weights initialization, fine-tuning, and retraining classification layers only, specifically utilizing optical coherence tomography images.