Abstract
Purpose :
Early diagnosis through clinical examination and fundus photography (FP) is of essential for age-related macular degeneration (AMD). Artificial intelligence (AI) has been utilized for computer vision applications, such as AMD diagnosis through FP; however, FP images in a database often vary in size and can pose a challenge due to laborious image preprocessing. This study proposes an AI model that distinguishes between normal retina, and wet and dry AMD through FP images and integrates a novel approach to handle discrepancies in image dimensions while upholding critical information necessary for diagnosis.
Methods :
Diagnosis and classification were achieved through two successive stages. Following normalization and selective data augmentation in image preprocessing, the SA Autoencoder model utilized convolutional, max pooling, reshaping, and concatenation layers in two branches to adapt varying input image dimensions to output dimensions of 224x224 px. A custom integrated loss function was utilized to maintain and enhance image integrity. Disease classification was accomplished through a BiT-M backbone followed by three repeated blocks of Group Normalization (GN), Weight Standardization (WS), and a dropout layer for optimized small batch training. A softmax function then assigned a predicted classification of normal retina, or wet or dry AMD. 648 FP images from The Comparisons of AMD Treatments Trials (CATT) sponsored by the University of Pennsylvania were allocated to training (70%), validating (15%), and testing (15%) the model, and the testing group was supplemented with a public ODiR dataset.
Results :
The proposed model outperformed state-of-the-art methods, achieving an accuracy, sensitivity, and specificity of 94%, 93.9%, and 97%, respectively, when trained on the private dataset. Remarkable metrics were attained on the OdiR public dataset with the highest accuracy, sensitivity, and specificity of 86%, 85.8%, and 93.5%, respectively.
Conclusions :
The proposed model utilized FP images to differentiate between healthy retina and wet and dry AMD. Varying input image dimensions were standardized to specific dimension by effectively retaining essential information and mitigating noise. It demonstrated promising results, recording among the highest performance metrics among other models using private and ODiR public databases.
This abstract was presented at the 2024 ARVO Annual Meeting, held in Seattle, WA, May 5-9, 2024.