Artificial intelligence-assisted diagnosis of glomerular nephritis using a pathological image analysis approach: a multicentre model development and validation study.

Background: The diagnosis of glomerular nephritis (GN) is based on kidney biopsy analysis by nephropathologists, a process that is manual, subjective, and labor-intensive, inevitably leading to interobserver variability. This study aims to develop and validate an artificial intelligence (AI)-assisted model for diagnosing GN using digital histological images from kidney biopsy. Methods: This multicentre model development and validation study included patients who underwent a kidney biopsy and were diagnosed with focal segmental glomerulosclerosis (FSGS), IgA nephropathy (IgAN), membranous nephropathy (MN), or minimal change disease (MCD). The study cohorts were derived from three institutions: Nanfang Hospital of Southern Medical University, Jinyu Diagnostic Center, and Huayin Diagnostic Center. The patients from Nanfang Hospital, enrolled between January 01, 2017, and December 31, 2021, were randomly divided into a training cohort (80%) and an internal validation cohort (20%). The patients from Jinyu Diagnostic Center and Huayin Diagnostic Center, enrolled from January 01, 2023, to December 31, 2023, and from January 01, 2017, to December 31, 2018, respectively, were used as the external validation cohorts. Patients with a diagnosis of one of the four specified glomerulonephritis types were included. The primary exclusion criterion was low-quality biopsy images. We utilised 106,988 glomeruli light microscopy images to develop an AI-assisted diagnostic model. The AI-assisted diagnostic model comprises three components: a glomerular localization module (GloSNet) for segmenting glomeruli, a feature extraction and fusion module for glomerular lesion, and a patient-level classification module to diagnose four-type (IgAN, MN, MCD, FSGS). Model performance was evaluated using F1-score, precision, recall and accuracy. Findings: The study included 6682 patients, of whom 1235 were in the training cohort (Nanfang Hospital), 312 in the internal validation cohort (Nanfang Hospital), 2483 in external validation cohort I (Jinyu Diagnostic Center), and 2652 in external validation cohort II (Huayin Diagnostic Center). The AI-assisted diagnostic model demonstrated strong performance, achieving F1-scores of 83.86% (95% confidence interval [CI]:81.64-86.03%), precision of 81.37% (95% CI: 79.14-83.67%), and recall of 87.84% (95% CI: 85.49-89.95%) in external validation cohort I, and F1-scores of 85.45% (95% CI: 83.66-87.29%), precision of 83.12% (95% CI: 81.26-85.03%), and recall of 88.94% (95% CI: 87.07-90.79%) in external validation cohort II. Interpretation: This study developed an AI-assisted diagnostic model to segment glomeruli, classify lesions, and diagnose four types of GN. By automating pathological diagnosing, the model may potentially be used to reduce clinicians' workload, and improve efficiency, ultimately supporting faster diagnoses. Despite demonstrating robust performance on independent datasets, the study's key limitations include its restriction to a Chinese population and four specific GN subtypes, a reliance on a single staining method, and challenges with model interpretability. Further studies will focus on model interpretation, incorporating multiple staining types, and including multiple GN. Funding: National Key R&D Program of China, the Key Technologies R&D Program of Guangdong Province, the Guangdong Provincial Clinical Research, the Program of Introducing Talents of Discipline to Universities, the National Natural Science Foundation of China, Eastern Scholars Program from Shanghai Municipal Education Commission and Clinical Research Initiation Plan of Southern Medical University.