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

Area of Science:

• Nephrology
• Medical Imaging
• Artificial Intelligence

Background:

• Transmission electron microscopy (TEM) is crucial for diagnosing renal diseases.
• Manual interpretation of TEM images is time-consuming and variable.
• Automated detection of ultrastructural features can improve diagnostic accuracy and efficiency.

Purpose of the Study:

• To introduce and evaluate deep learning architectures based on YOLOv8-OBB for automated detection of six key ultrastructural features in kidney biopsy TEM images.
• To refine the YOLOv8-OBB architecture for enhanced detection of small, low-contrast, and variably oriented features.
• To assess the clinical applicability of these deep learning tools.

Main Methods:

• A modified YOLOv8-OBB architecture was developed with a grayscale input channel, a high-resolution P2 feature pyramid with refinement blocks (FPRbl), and a four-branch oriented detection head.
• Two pretrained variants were compared: a previous model (Pretrained + GSch + 4FExL) and a newly developed model (Pretrained + FPRbl).
• Performance was evaluated using quantitative metrics (F1-score, mAP@0.5) and qualitative expert visual inspection.

Main Results:

• Both models demonstrated strong performance, with F1-scores of 0.93 and 0.92, and mAP@0.5 scores of 0.953 and 0.941, respectively.
• The (Pretrained + GSch + 4FExL) model showed higher recall for subtle findings.
• The (Pretrained + FPRbl) model produced cleaner, higher-confidence bounding boxes.

Conclusions:

• Targeted architectural refinements in YOLOv8-OBB effectively enhance the detection of challenging ultrastructural features in renal TEM images.
• The developed deep learning models show clinical applicability for improving diagnostic efficiency and reducing interpretive variability in kidney pathology.
• Translation into a web-based platform (Renal-AI) demonstrates the practical utility of these AI tools.