In June 2025, the team led by Prof. Bin Sheng from the Ministry of Education Key Laboratory of Artificial Intelligence, School of Computer Science, Shanghai Jiao Tong University, together with Prof. Rongping Dai’s team from the Department of Ophthalmology, Peking Union Medical College Hospital / Key Laboratory of Fundus Diseases, Chinese Academy of Medical Sciences, published a study titled “A deep learning system for detecting systemic lupus erythematosus from retinal images” in Cell Reports Medicine. 

The work reports the world’s first deep-learning–based screening system for systemic lupus erythematosus (SLE) using retinal images—DeepSLE. With a single retinal photograph, DeepSLE can noninvasively identify SLE and its common retinal and renal complications, offering a novel technological pathway to improve early-screening efficiency and equity for rare diseases. The study was selected as the cover article of Cell Reports Medicine.

Figure legend: DeepSLE selected as the cover story of the July 2025 issue of Cell Reports Medicine.

Research Overview

Systemic lupus erythematosus (SLE) is a multi-organ autoimmune disease primarily affecting women of childbearing age. Its core pathology involves immune dysfunction–driven vascular inflammation, which can affect the kidneys, retina, and nervous system. Because SLE symptoms are often insidious and nonspecific, patients worldwide frequently encounter late diagnosis and substantial organ damage. This challenge is especially pronounced in primary healthcare settings, where limited rheumatology expertise and costly laboratory testing render early screening almost inaccessible.

The retina is the only part of the human body where microvasculature can be directly and noninvasively visualized, and retinal vascular abnormalities can mirror systemic pathological processes. Previous studies have demonstrated strong associations between retinal characteristics and diabetes, as well as cardiovascular and cerebrovascular diseases. In SLE, retinal abnormalities are also common, and microvascular morphology correlates strongly with disease activity (e.g., SLEDAI score). DeepSLE is built upon the physiological hypothesis of an “eye–systemic vascular damage association model,” advancing retinal imaging from a traditional ophthalmic diagnostic tool to a “health sentinel” for systemic diseases.

As a rare disease with low prevalence, SLE suffers from severe data scarcity. To address this “small-sample trap”, the research team adopted a combined strategy of self-supervised pretraining followed by multi-task fine-tuning. The model was first pretrained in an unsupervised manner on over 660,000 retinal fundus images from the general population to learn robust visual representations. It was then fine-tuned on specific asks including SLE, lupus retinopathy (LR), and lupus nephritis (LN), significantly enhancing its ability to identify rare-disease features. To further improve generalization across disease stages, the team introduced a curriculum learning strategy that mimics clinicians’ diagnostic reasoning—from easily distinguishable samples to progressively complex cases—thus gradually strengthening the model’s decision-making capabilities. Across a multi-ethnic validation dataset containing 247,718 images from China and the United Kingdom, DeepSLE achieved an AUROC of 0.822–0.969 for SLE identification and demonstrated strong performance in detecting LR and LN. The system remained robust across subgroup analyses stratified by sex, age, ethnicity, and socioeconomic status.