A Disease-Specific Foundation Model Using Over 100K Fundus Images: Release and Validation for Abnormality and Multi-Disease Classification on Downstream Tasks

TL;DR

This work addresses the need for robust, scalable fundus image analysis by introducing disease-specific foundation models trained on a large fundus dataset. It compares two upstream pretraining strategies—a fundus-only model and a two-step ImageNet+Fundus approach—across abnormality classification, multi-disease classification, and vessel segmentation, using high-resolution inputs and rigorous evaluation including five-fold cross-validation and external validation. The ImageNet+Fundus pretraining consistently yields the strongest performance on abnormality and multi-disease tasks, demonstrates strong generalization to JSIEC and RFMiD datasets, and remains effective under data-scarce conditions, while vessel segmentation highlights limitations in disease-focused representations for structural tasks. The work provides open-source code and pretrained weights, illustrating practical utility for rapid abnormality screening and paving the way for broader retinal-structure analysis and downstream ophthalmic applications.

Abstract

Artificial intelligence applied to retinal images offers significant potential for recognizing signs and symptoms of retinal conditions and expediting the diagnosis of eye diseases and systemic disorders. However, developing generalized artificial intelligence models for medical data often requires a large number of labeled images representing various disease signs, and most models are typically task-specific, focusing on major retinal diseases. In this study, we developed a Fundus-Specific Pretrained Model (Image+Fundus), a supervised artificial intelligence model trained to detect abnormalities in fundus images. A total of 57,803 images were used to develop this pretrained model, which achieved superior performance across various downstream tasks, indicating that our proposed model outperforms other general methods. Our Image+Fundus model offers a generalized approach to improve model performance while reducing the number of labeled datasets required. Additionally, it provides more disease-specific insights into fundus images, with visualizations generated by our model. These disease-specific foundation models are invaluable in enhancing the performance and efficiency of deep learning models in the field of fundus imaging.

Figures (4)

• Figure 1: Illustrates the application of a pretrained fundus-specific model in processing and enhancing fundus images.
• Figure 2: Comprehensive analysis of model performance across different resolutions and data fractions. (A) mean AUC scores for three pre-trained models at varying image resolution within the LF method. (B) visual representations of the activation map using Grad-CAM for the ImageNet + Fundus pre-trained model, showing areas of interest that influence the model’s predictions for the class of fundus abnormalities. (C) mean AUC scores across difference data fractions for models trained from Scratch, and those pre-trained with ImageNet, Fundus, and ImageNet + Fundus weights.
• Figure 3: t-SNE visualization of multi-disease classification results with foundation model of ImageNet + Fundus.
• Figure 4: Performance evaluation across data fractions for vessel segmentation under the full fine-tuning method.