Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

RET-CLIP: A Retinal Image Foundation Model Pre-trained with Clinical Diagnostic Reports

Jiawei Du, Jia Guo, Weihang Zhang, Shengzhu Yang, Hanruo Liu, Huiqi Li, Ningli Wang

TL;DR

RET-CLIP introduces a retinal vision-language foundation by jointly pretraining on binocular color fundus photographs and clinical diagnostic reports, including Chinese text. The approach decouples left-eye, right-eye, and patient-level information and uses monocular- and patient-level contrastive losses to align image and text representations, resulting in a robust, generalizable feature space. Evaluated across eight datasets in four diagnostic categories, RET-CLIP achieves state-of-the-art performance in both linear probing and fine-tuning, demonstrating strong cross-domain transfer for ophthalmic tasks. The work highlights the value of integrating rich clinical text into medical CLIP-style pretraining and provides publicly available code and pretrained weights for replication.

Abstract

The Vision-Language Foundation model is increasingly investigated in the fields of computer vision and natural language processing, yet its exploration in ophthalmology and broader medical applications remains limited. The challenge is the lack of labeled data for the training of foundation model. To handle this issue, a CLIP-style retinal image foundation model is developed in this paper. Our foundation model, RET-CLIP, is specifically trained on a dataset of 193,865 patients to extract general features of color fundus photographs (CFPs), employing a tripartite optimization strategy to focus on left eye, right eye, and patient level to reflect real-world clinical scenarios. Extensive experiments demonstrate that RET-CLIP outperforms existing benchmarks across eight diverse datasets spanning four critical diagnostic categories: diabetic retinopathy, glaucoma, multiple disease diagnosis, and multi-label classification of multiple diseases, which demonstrate the performance and generality of our foundation model. The sourse code and pre-trained model are available at https://github.com/sStonemason/RET-CLIP.

RET-CLIP: A Retinal Image Foundation Model Pre-trained with Clinical Diagnostic Reports

TL;DR

RET-CLIP introduces a retinal vision-language foundation by jointly pretraining on binocular color fundus photographs and clinical diagnostic reports, including Chinese text. The approach decouples left-eye, right-eye, and patient-level information and uses monocular- and patient-level contrastive losses to align image and text representations, resulting in a robust, generalizable feature space. Evaluated across eight datasets in four diagnostic categories, RET-CLIP achieves state-of-the-art performance in both linear probing and fine-tuning, demonstrating strong cross-domain transfer for ophthalmic tasks. The work highlights the value of integrating rich clinical text into medical CLIP-style pretraining and provides publicly available code and pretrained weights for replication.

Abstract

The Vision-Language Foundation model is increasingly investigated in the fields of computer vision and natural language processing, yet its exploration in ophthalmology and broader medical applications remains limited. The challenge is the lack of labeled data for the training of foundation model. To handle this issue, a CLIP-style retinal image foundation model is developed in this paper. Our foundation model, RET-CLIP, is specifically trained on a dataset of 193,865 patients to extract general features of color fundus photographs (CFPs), employing a tripartite optimization strategy to focus on left eye, right eye, and patient level to reflect real-world clinical scenarios. Extensive experiments demonstrate that RET-CLIP outperforms existing benchmarks across eight diverse datasets spanning four critical diagnostic categories: diabetic retinopathy, glaucoma, multiple disease diagnosis, and multi-label classification of multiple diseases, which demonstrate the performance and generality of our foundation model. The sourse code and pre-trained model are available at https://github.com/sStonemason/RET-CLIP.
Paper Structure (14 sections, 7 equations, 1 figure, 5 tables)

This paper contains 14 sections, 7 equations, 1 figure, 5 tables.

Table of Contents

  1. Introduction
  2. Method
  3. Data Collection and Preprocessing
  4. Model Architecture
  5. Training Objective
  6. Implementation
  7. Experiments
  8. Tasks and Datasets
  9. Comparision Methods and Evaluation Metrics
  10. Result
  11. Ablation study
  12. Conclusion
  13. Acknowledgments.
  14. Disclosure of Interests.

Figures (1)

  • Figure 1: Overview of the RET-CLIP foundation model.