Abstract
The diagnosis of membranous nephropathy (MN) has been reliant on the identification of glomerular basement membrane (GBM) variations and lesions at both macro and micro levels. At the macro level, light microscopy (LM) has been used to reveal spike-like projections that indicate pathological changes, whereas at the micro level, transmission electron microscopy (TEM) has been employed to identify GBM thickening. However, qualitative diagnosis has been limited by inter-pathologist variability, creating the need for deep learning approaches capable of quantifying pathological changes and predicting MN progression. In this study, an AI-driven framework based on the Mamba model has been proposed, in which the area and proportion of spike-like projections are quantified at the macro level, and GBM thickness is segmented and measured at the micro level. Classical machine learning models are then applied to predict MN progression based on pathological indicators extracted through factor analysis. Unlike prior approaches, the framework has been designed to emulate the diagnostic workflow of pathologists by integrating LM and TEM images for joint analysis. Experiments on an external dataset of 109 cases demonstrated strong performance in glomeruli classification, GBM segmentation, and MN progression prediction. These findings highlight the potential of multi-scale integrated quantification to provide objective, reproducible, and clinically interpretable assessment of MN progression.