Renal-AI: A Deep Learning Platform for Multi-Scale Detection of Renal Ultrastructural Features in Electron Microscopy Images

TL;DR

This paper introduces Renal-AI, a deep learning platform that automates the detection of kidney ultrastructural features in electron microscopy images, improving diagnostic accuracy and efficiency.

Contribution

The study introduces architectural refinements to YOLOv8-OBB for detecting small, low-contrast renal ultrastructural features in TEM images.

Findings

A modified YOLOv8-OBB model achieved an F1-score of 0.93 and mAP@0.5 of 0.953 for detecting renal ultrastructural features.

The new model (Pretrained + FPRbl) achieved an F1-score of 0.92 and mAP@0.5 of 0.941, showing strong performance.

Expert evaluation showed complementary strengths between the two models in detecting subtle and frequent features.

Abstract

Background/Objectives: Transmission electron microscopy (TEM) is an essential tool for diagnosing renal diseases. It produces high-resolution visualization of glomerular and mesangial ultrastructural features. However, manual interpretation of TEM images is labor-intensive and prone to interobserver variability. In this study, we introduced and evaluated deep learning architectures based on YOLOv8-OBB for automated detection of six ultrastructural features in kidney biopsy TEM images: glomerular basement membrane, mesangial folds, mesangial deposits, normal podocytes, podocytopathy, and subepithelial deposits. Methods: Building on our previous work, we propose a modified YOLOv8-OBB architecture that incorporates three major refinements: a grayscale input channel, a high-resolution P2 feature pyramid with refinement blocks (FPRbl), and a four-branch oriented detection head designed to detect small-to-large structures at multiple image scales (feature-map strides of 4, 8, 16, and 32 pixels). We compared two pretrained variants: our previous YOLOv8-OBB model developed with a grayscale input channel (GSch) and four additional feature-extraction layers (4FExL) (Pretrained + GSch + 4FExL) and the newly developed (Pretrained + FPRbl). Results: Quantitative assessment showed that our previously developed model (Pretrained + GSch + 4FExL) achieved an F1-score of 0.93 and mAP@0.5 of 0.953, while the (Pretrained + FPRbl) model developed in this study achieved an F1-score of 0.92 and mAP@0.5 of 0.941, demonstrating strong and clinically meaningful performance for both approaches. Qualitative assessment based on expert visual inspection of predicted bounding boxes revealed complementary strengths: (Pretrained + GSch + 4FExL) exhibited higher recall for subtle or infrequent findings, whereas (Pretrained + FPRbl) produced cleaner bounding boxes with higher-confidence predictions. Conclusions: This study presents how targeted architectural refinements in YOLOv8-OBB can enhance the detection of small, low-contrast, and variably oriented ultrastructural features in renal TEM images. Evaluating these refinements and translating them into a web-based platform (Renal-AI) showed the clinical applicability of deep learning-based tools for improving diagnostic efficiency and reducing interpretive variability in kidney pathology.

Figures