Two-Stage Deep Learning Networks for Diagnosing and Staging Membranous Glomerulonephritis From Electron Microscopy Images.

PURPOSE: The purpose of this study was to use transmission electron miscroscopy images to develop a deep learning-based artificial intelligence model to detect and stage membranous glomerulonephritis (MN), a major cause of nephrotic syndrome in adults.

MATERIALS AND

METHODS

A comprehensive data set of micrographs was used to train and test the proposed model. The architecture consisted of a 2-stage model based on vision transformer networks. The first stage detected membrane regions independent of disease stage; the second stage classified each region according to the stage of MN; and the final image-level classification was determined by a majority vote among the individual membrane regions. Model generalization was evaluated using an independent data set containing patient micrographs, acquired using a microscope and camera, and evaluated by different practitioners.

RESULTS

The proposed model achieved an accuracy of 0.9231, sensitivity of 0.9240, and specificity of 0.9167. The model maintained strong performance on the external data set (accuracy, 0.8905).

CONCLUSIONS

The proposed 2-stage model enhances accuracy, robustness, and clinical interpretability in MN staging from transmission electron microscopy images, thus showing strong potential for integration into routine diagnostic workflows. Article History.