Medical Artificial Intelligence · asia
Making Two Types of Retinal Imaging Complement Each Other: RETBind Improves Performance Across Nine Ophthalmic AI Benchmarks
By learning from paired OCT and ultra-widefield fundus images from the same patient, RETBind aims to help AI interpret complementary retinal clues. Results from nine datasets support this research approach, but clinical validation is still needed to demonstrate benefits in clinical practice.
Different examinations of the same eyes reveal different clues. Optical coherence tomography (OCT) shows the layered structure of the retina, while ultra-widefield fundus photography provides a broader view of the fundus. Connecting this information through artificial intelligence is a challenge in ophthalmic imaging research; the new RETBind model takes the approach of learning relationships between two types of images from the same patient.
The study was published on November 3 in *npj Digital Medicine*, with participation from teams at South Korea’s Sungkyunkwan University and ophthalmic medical institutions, among others. Sungkyunkwan University’s Data Science and Artificial Intelligence Laboratory had announced in October that a paper with the same title had been accepted, listing participating researchers including Jian, Hyunseon, JeeEun, and Sihyeon. That announcement confirmed the study’s progress but provided no additional clinical validation results.
RETBind is a foundation model that can be used for subsequent imaging tasks. The study pretrained the model using OCT and ultra-widefield fundus images paired at the patient level, with both image types sharing a single encoder that converts images into features the model can recognize. The focus of this design is to incorporate complementary information from different examinations into joint learning, rather than processing each image type separately.
The research team reported that RETBind improved performance across all nine benchmark datasets covering eye disease classification and image segmentation. Classification tasks focus on identifying disease categories, while segmentation tasks require the model to mark specific regions within images. The authors highlighted improvements in the area under the precision–recall curve (AUPR) when sample sizes were uneven across classes. This helps assess performance in identifying minority classes, but it cannot be translated directly into a reduction in missed diagnoses in clinical practice.
The nine benchmarks provide evidence that the model can be used for different tasks, but they have yet to establish prospective clinical benefits. Before use in actual care, the model’s reliability across different hospitals, devices, and patient populations must still be confirmed, along with whether incorporating it into physicians’ image interpretation workflows improves diagnosis, referral, or subsequent management. The current results are also insufficient to conclude that the model is ready for independent diagnostic use.
The paper also used exploratory analyses to examine the possibility of predicting systemic diseases from retinal images; this component should be viewed as a lead for further research. RETBind’s clearer contribution at present is to provide a testable technical approach for integrating multiple types of ophthalmic images from the same patient. Whether its benchmark performance can translate into patient benefits remains a question to be answered in real-world care workflows.