Abstract
Purpose :
Deep learning (DL) models traditionally necessitate extensive datasets for effective training and fine-tuning. RetFound, a novel foundational model tailored for retinal images, demonstrates significant promise as a pre-training tool. However, it is uncertain if RetFound maintain its performance when fine-tuned with a considerably smaller dataset. Therefore, our study investigates RetFound's adaptability and efficiency in such contexts, aiming to transform the development of AI models for pathology-based visual impairment (VI) detection.
Methods :
This study introduces two deep learning models for detecting pahtology VI using retinal images: the Swin Vision Transformer (VT) and a VT model pretrained on RetFound. Training involved 13,170 images from the Singapore Epidemiology of Eye Disease (SEED) study, with an additional 3,803 images for testing. The adaptability of RetFound was further evaluated on smaller subsets of the SEED dataset, comprising 10% (1,317 images), and a targeted training on 100 normal and 100 case images. External validation was conducted using datasets from three external population studies. Pathology VI was defined as presence of eye diseases with best-corrected visual acuity worse than 6/12. We evaluated the models' performance using Area under the curve (AUC) scores.
Results :
The Swin VT model achieved an AUC of 0.951 on the internal test set, with AUCs ranging from 0.935 to 0.953 across external test sets. The RetFound model, pretrained and fine-tuned on the full training set, showed a comparable internal test set AUC of 0.947, with external test set AUCs between 0.919 to 0.956. Interestingly, pretraining on just 10% of the original dataset resulted in only a marginal AUC decrease to 0.93 internally and 0.861 to 0.954 externally. Training on 100 cases and 100 normals led to an internal AUC of 0.939, with external test set performance marginally reduced to between 0.883 and 0.960.
Conclusions :
Pre-training on RetFound and fine-tuning on smaller datasets can produce high-performance models for pathology VI detection, comparable to those trained on larger datasets. This approach underscores the potential for developing efficient, high-performance models for ocular disease detection in ophthalmology.
This abstract was presented at the 2024 ARVO Annual Meeting, held in Seattle, WA, May 5-9, 2024.