Original Investigation 

Nephrology

AI-Based System for Analysis of Electron Microscope Images in Glomerular Disease

Pengcheng Ma, Jinbang Li, Zhengyu Zhang, et al

JAMA Netw Open 2025;8;(10):e2534985. doi:10.1001/jamanetworkopen.2025.34985

Key Points

Question  What is the performance of an artificial intelligence (AI) system for the analysis of kidney transmission electron microscopy (TEM) images for diagnosing glomerular disease?

Findings  This diagnostic study including 160 727 images from 31 670 patients across 6 health centers describes the development and validation of TEM-AID, an AI-based system to diagnose glomerular disease. TEM-AID accurately segmented glomerular structures and directly predicted glomerulonephritis subtypes, achieving high internal diagnostic performance and demonstrating consistent external validation across 5 test sets; additionally, in a comparative study involving 454 patients, TEM-AID outperformed pathologists.

Meaning  This diagnostic study found that TEM-AID significantly enhanced diagnostic efficiency and accuracy in kidney pathology by providing quantitative TEM analysis tools, supporting pathologists in clinical practice.

Abstract

Importance  Kidney biopsy pathology via transmission electron microscopy (TEM) is essential for diagnosing glomerular diseases, offering critical information on glomerular basement membrane (GBM) thickness, foot process (FP) number, and electron-dense deposits (EDDs). These tasks are laborious and time-consuming.

Objective  To develop and validate an artificial intelligence (AI) diagnostic system, TEM image–based AI-assisted device (TEM-AID), that accurately segments and measures glomerular ultrastructures (including the GBM, FPs, and EDDs) and determines glomerular disease subtypes using TEM images.

Design, Setting, and Participants  This diagnostic study used a large, multicenter cohort including 160 727 TEM images from 31 670 patients with chronic kidney disease across 6 medical centers from January 2021 to December 2023. TEM-AID was trained and validated on 26 650 patients from 1 center and tested externally on 5020 patients (5 test sets) plus a human-AI test set (454 patients representing 7 glomerular disease subtypes). Data were analyzed from January to December 2024.

Exposures  TEM-AID integrates 4 modules. Segmentation combined YOLO-v8 detection, segment anything model, and human-in-the-loop refinement to segment GBMs, podocyte FPs, and EDDs. Measurement quantified GBM thickness, FP fusion degree, and EDD deposition sites. Classification used least absolute shrinkage and selection operator–selected deep learning and statistical features with a stacking classifier to diagnose 7 glomerular disease subtypes: immunoglobin A nephropathy, membranous nephropathy, lupus nephritis, diabetic nephropathy, minimal change disease, mesangial proliferative glomerulonephritis, and thin basement membrane nephropathy.

Main Outcomes and Measures  Outcomes of interest were segmentation performance (mean intersection-over-union [IOU], Dice coefficient), subtype classification accuracy, area under the receiver operating characteristic curve (AUC), and human-AI diagnostic concordance.

Results  A total of 31 670 patients (mean [SD] age, 43.2 [16.5] years; 17 372 [54.9%] male) contributed 160 727 TEM images for analysis. Segmentation achieved a mean (SD) IOU of 0.835 (0.062) and Dice of 0.874 (0.023). Subtype classification accuracy was 0.911 (95% CI, 0.904-0.918) in internal validation and 0.895 to 0.914 in external tests. Macro-AUC ranged from 0.972 to 0.989 across cohorts. In human-AI testing (454 patients), TEM-AID accuracy (0.886 (95% CI, 0.859-0.912]; AUC, 0.963 [95% CI, 0.937-0.989]) exceeded clinicians’ unaided performance. Clinicians’ accuracy improved by a mean (SD) of 11.7% (5.2%) when they used TEM-AID.

Conclusions and Relevance  In this multicenter diagnostic study, TEM-AID precisely quantified glomerular ultrastructures and determined glomerular disease subtypes from TEM images, significantly enhancing diagnostic efficiency and accuracy. This system provides quantitative evaluation tools to support clinical pathologists in diagnostic workflows, demonstrating robust multicenter performance.