Improving Diagnostic Accuracy in Glomerular Diseases with AI-Based GBM Thickness Assessment

Human glomerular diseases are pathological conditions that affect the renal corpuscles causing morphological and structural alterations, including the thinning and thickening of glomerular basement membranes (GBMs) [1]. Changes in the GBM thickness can be exclusively recognized at the ultrastructural level for its highest magnifications and resolution. Transmission electron microscopy (TEM) is a pivotal technique for the GBM thickness measurements and a diagnostic criterion in some glomerular diseases [2-3]. Without standardized procedures, these measurements are manually performed on TEM images by pathologists requiring time, high experience and strategies to make them accurate and reproducible; therefore, the subjectivity and operator-dependent expertise could affect the measurement and the related final diagnosis. In this study, we show the application of a fully automated pipeline for the GBM segmentation and thickness estimation, using ultrastructural images at different magnifications, contrasts, and described by complex and irregular geometrical shapes. The pipeline implements a convolutional neural network model trained using an active semi-supervised learning training strategy, which guarantees an easier extension to novel datasets. The model was trained to identify and segment GBMs, validating its accuracy against manual annotation images. The obtained information was used by a dedicated image processing algorithm for the GBM thickness evaluation. The accuracy of the developed pipeline was validated comparing the predicted measurements with the diagnostic reports obtained by manual assessment of TEM images, stratifying the GBM thicknesses in standard categories (normal, thin, thick). This pipeline was able to correctly classify membrane categories with an accuracy of more than 75%, reproducing the manual thickness estimations with a correlation of 0.85 (Pearson’s R2). The proposed fully automated solution could represent an important tool for clinical applications, providing a viable clinical decision support system for diagnosis. The automation of the procedure could assist clinicians speeding-up routine procedures, improving diagnostic accuracy and avoiding human subjectivity issues. References [1] E.N. Pavlisko and D.N. Howell, “The continued vital role of electron microscopy in the diagnosis of renal disease/dysfunction”, Ultrastruct Pathol, pp. 1-8, 2013. [2] M. Yamashita, M.Y. Lin, J. Hou, K.Y.M Ren and M. Haas, "The continuing need for electron microscopy in examination of medical renal biopsies: examples in practice", Glomerular Dis, pp. 145-59, 2021. [3] F.E. Dische, "Measurement of glomerular basement membrane thickness and its application to the diagnosis of thin-membrane nephropathy", Arch Pathol Lab Med, pp. 43-9, 1992.