eClinicalMedicine | AI-assisted diagnosis of glomerular nephritis

This study developed and validated an artificial intelligence (AI)-assisted diagnostic model for automated glomerular nephritis (GN) diagnosis based on kidney biopsy, demonstrating high accuracy and stability. Validated through multicenter data, the article shows the AI model's robust performance in classifying four GN types (IgA nephropathy, membranous nephropathy, focal segmental glomerulosclerosis, and minimal change disease), providing an essential tool to enhance pathologists' diagnostic efficiency and consistency.

 

Literature Overview
This article titled 'Artificial intelligence-assisted diagnosis of glomerular nephritis using a pathological image analysis approach: a multicentre model development and validation study', published in eClinicalMedicine, reviews the model development and validation process for AI-assisted GN diagnosis. Through multicenter research utilizing 106,988 glomerular microscopic images, the study constructed an integrated model comprising three components: glomerular localization, lesion feature extraction, and patient-level classification, achieving overall performance metrics with F1-scores exceeding 85%. The research also explored AI models combining light microscopy with immunofluorescence to further improve diagnostic accuracy.

Background Knowledge
Glomerular nephritis (GN) represents a group of diseases characterized by glomerular inflammation, serving as a major cause of end-stage kidney disease (ESKD), particularly in developing countries. Currently, GN diagnosis relies on histopathological analysis of kidney biopsies, a process requiring experienced pathologists to manually identify glomerular lesions but suffering from subjectivity, time consumption, and poor reproducibility. While AI applications in digital pathology image analysis have shown exceptional capabilities in breast and lung cancer diagnosis, a unified and efficient framework for GN diagnosis remains lacking. This study pioneers AI application in GN pathology, establishing an automated, reproducible diagnostic tool to address manual slide review limitations. Although previous studies explored AI potential in glomerular segmentation and lesion detection, they lacked large-scale multicenter data support. Through training on over 100,000 images, this research validated the AI model's stability in real-world datasets while identifying limitations related to race, disease subtypes, and multichromatic methods, providing directions for future research.

 

 

Research Methods and Experiments
This multicenter study collected pathological image data from 6,682 kidney biopsy patients, including 1,235 cases for training, 312 for internal validation, and 2,483/2,652 as two external validation cohorts. All images underwent quality control and standardization, retaining only Periodic Acid-Silver Methenamine (PASM)-stained light microscopy images. The AI model comprises three core modules: glomerular localization module (GloSNet) for image segmentation, feature fusion module for extracting and integrating glomerular lesion features, and patient-level classification module for final diagnosis. Evaluation metrics included F1-score, precision, recall, and accuracy. Additionally, the study explored a dual-branch model incorporating immunofluorescence (IF) images to further enhance diagnostic accuracy.

Key Conclusions and Perspectives

• The AI model achieved 83.86% F1-score in external validation cohort 1 and 85.45% in cohort 2, demonstrating high diagnostic consistency across different datasets.
• The glomerular segmentation module (GloSNet) showed excellent performance at pixel and glomerular levels with F1-scores of 93.01% and 96.10% respectively, indicating high accuracy in automatic glomerular region identification.
• The feature fusion strategy based on kernel density estimation effectively integrated different lesion features, enhancing model performance in multi-category GN classification tasks - particularly achieving 97.13% F1-score in MN (membranous nephropathy).
• The dual-branch AI model incorporating immunofluorescence data improved four-class F1-score to 91.00%, demonstrating IF data's effectiveness in enhancing AI diagnostic capabilities.
• Through visualized attention heatmaps, the research team confirmed the AI model's ability to precisely identify GN-related pathological features such as MN's basement membrane thickening and FSGS's sclerosis regions.

Research Significance and Prospects
This study demonstrates AI's high accuracy in glomerular nephritis diagnosis for the first time, providing pathologists with an objective, reproducible auxiliary tool. Future research should expand to more races and GN subtypes while optimizing model interpretability to improve clinical applicability. The research team also recommends integrating multiple staining methods and multi-omics data to enhance model generalization capabilities for broader applications across global medical centers.

 

 

Conclusion
This study successfully developed an AI-assisted glomerular nephritis diagnostic model based on kidney biopsy light microscopy images, achieving efficient classification of four GN types (IgAN, MN, FSGS, MCD). The model demonstrated excellent performance in both external validation cohorts with F1-scores exceeding 83%. Visualized analysis confirmed the AI model's ability to identify GN-related pathological features. While currently limited by race and disease subtype specificity, this tool significantly reduces pathologists' workload and improves diagnostic consistency. Future directions include expanding to more disease subtypes, integrating multiple staining methods, enhancing model interpretability, and exploring real-time diagnostic and telemedicine applications.

 

Sheng Nie, Nan Jia, Haobo Chen, Qi Zhang, and Fan Fan Hou. Artificial intelligence-assisted diagnosis of glomerular nephritis using a pathological image analysis approach: a multicentre model development and validation study. eClinicalMedicine.