DeepDKD: New Breakthrough in Diagnosing Diabetic Kidney Disease Using Retinal Images

Professor Sheng Bin's team from the Department of Computer Science and Technology at Shanghai Jiao Tong University, in collaboration with Professor Tianyin Huang’s team from Tsinghua University, has developed a novel deep learning system called DeepDKD for non-invasive diagnosis of diabetic kidney disease (DKD). This groundbreaking research was published in the prestigious journal Lancet Digital Health in May 2025. The DeepDKD system uses retinal images to screen for DKD and accurately distinguish between diabetes-related nephropathy (DN) and non-diabetic kidney disease (NDKD), offering a significant advancement in the field. DKD, a serious microvascular complication of diabetes, is a leading cause of end-stage renal disease worldwide and imposes a substantial burden on global health and economies. Approximately 40% of diabetes patients are at risk of developing DKD. The increasing prevalence of this condition is linked to longer diabetes duration, poor glycemic control, and an aging population. According to recent studies, the number of disability-adjusted life years (DALYs) attributed to DKD reached over 15 million in 2021, and it is projected to rise to 19 million by 2030, particularly affecting countries with medium sociodemographic indices (SDIs). Moreover, DKD is strongly associated with higher cardiovascular mortality, creating a detrimental "diabetes-kidney disease-cardiovascular events" cycle. Current screening methods for DKD primarily rely on estimated glomerular filtration rate (eGFR) and urine albumin-to-creatinine ratio (ACR). These methods face several challenges: 1. Low Patient Awareness: Early symptoms of DKD are often subtle and easily overlooked by patients. 2. High Variability in Urine ACR: While urine samples are relatively easy to collect, the biological variability in albuminuria can range from 20% to 50%, complicating accurate screening. 3. Difficulty in Differentiating DN and NDKD: Kidney biopsy, though considered the gold standard for diagnosing DN, is invasive and costly, leading to low patient acceptance. Furthermore, treatment protocols for DN do not significantly change based on biopsy results, rendering routine biopsies unnecessary for all DKD patients. The research team identified the retina as a potential solution, being the only microvascular system in the body that can be directly observed. They developed DeepDKD, which processes retinal images using large-scale weakly supervised momentum contrast learning methods to extract transferable features. This system successfully identifies the stages of DKD and distinguishes between DN and NDKD. To validate DeepDKD, the researchers analyzed a diverse dataset comprising over 700,000 retinal images from multiple racial groups across China, Singapore, the UK, Malaysia, and Australia. The results demonstrated the algorithm's robustness and generalizability across different populations. In a prospective real-world study, DeepDKD showed higher sensitivity in screening for DKD compared to clinical data models. Additionally, patients identified by DeepDKD as having DN or NDKD exhibited significant differences in kidney function outcomes (eGFR) over a 4.6-year follow-up period, indicating that the system can precisely differentiate between the two conditions. This technological innovation has the potential to revolutionize DKD screening and management, making it non-invasive, efficient, and accessible. DeepDKD can simultaneously screen for diabetic retinopathy (DR) and DKD during a single retinal imaging session, aligning well with the needs of primary healthcare settings. By providing a non-invasive differential diagnosis, DeepDKD addresses the limitations of traditional kidney biopsies, reducing the need for invasive procedures and offering ethical and practical solutions for hospitals with limited resources. Looking ahead, integrating DeepDKD into primary care could lead to earlier detection and intervention for DKD, potentially slowing its progression to end-stage renal disease and significantly reducing the strain on healthcare systems. The team plans to continue refining and clinically implementing DeepDKD, aiming to establish a new era of DKD management characterized by "non-invasiveness, precision, and accessibility." In this study, Professor Tianyin Huang, Professor Sheng Bin, and Professor Hua Ting Li are the corresponding authors. The primary authors include doctoral students Ziyao Meng (supervised by Professor Sheng Bin), Zhouyu Guan (supervised by Professor Weiping Jia), and Shujie Yu (supervised by Professor Hua Ting Li), along with medical professionals and graduate students from both universities and international institutions. Support was provided by various organizations, including the National Basic Diabetes Prevention and Management Office, Huazhong University of Science and Technology, Peking Union Medical College, Chinese University of Hong Kong, and multidisciplinary teams from the UK, USA, Singapore, Australia, and Malaysia. The research was funded by the National Key R&D Program, National Natural Science Foundation, Beijing Natural Science Foundation, and Shanghai Science and Technology Commission, among others. Computational resources were supported by Shanghai Jiao Tong University's "Jiao Wesuan" platform and the AI for Science open-source data platform.