Confidence-guided multi-image fusion boosts diabetic retinopathy screening accuracy by 12%

A team of researchers (Ananya Raghu, Anisha Raghu, Alice S. Tang, Yannis M. Paulus, Tyson N. Kim, Tomiko T. Oskotsky) has introduced a novel framework for diabetic retinopathy screening that fuses multiple fundus images guided by model confidence. Unlike traditional cascaded pipelines that rely on human-annotated image quality labels, the new method uses confidence scores to identify unreliable predictions and prompt retakes when needed. Tested on two datasets (mBRSET with 1,234 patients and BRSET with 7,599 patients) using a RETFoundGreen backbone, the approach achieves 91% balanced accuracy on mBRSET and 97% on BRSET at 70% coverage – roughly 12% and 6% improvements respectively over standard cascade filtering. Sensitivity at 50% coverage reaches 94% and 96%, versus 61% and 86% for the cascade pipeline.

The key insight is that human-annotated quality labels are weakly associated with actual diagnostic performance, while confidence-based filtering consistently outperforms them. The framework requires only a single inference pass per image and uses a lightweight backbone, making it compatible with low-latency mobile screening systems in low- and middle-income countries where access to specialist care is limited. By aggregating multiple retinal views, patients receive more reliable predictions, reducing incorrect diagnoses during early screening. This work, published on arXiv (2607.03643), represents a practical step toward AI-assisted eye care in resource-constrained settings.