Assessment of glomerular morphological patterns by deep learning algorithms.

Area of Science:

• Nephropathology
• Artificial Intelligence
• Digital Pathology

Background:

• Renal pathology diagnosis relies on identifying diverse morphological lesion signatures.
• Artificial intelligence (AI) has shown potential in recognizing clear-cut glomerular structures.
• This study addresses AI's capacity for complex glomerular changes in diagnostic dilemmas.

Purpose of the Study:

• To evaluate deep learning algorithms for recognizing complex glomerular structural changes.
• To assess AI's ability to handle diagnostic challenges in nephropathology.
• To develop a reliable AI tool for identifying discrete and overlapping morphological changes.

Main Methods:

• Defined nine classes of glomerular morphological patterns.
• Trained twelve convolutional neuronal network (CNN) models using two datasets (12,253 and 11,142 images).
• Utilized expert-defined and consensus datasets for robust model training.

Main Results:

• CNN models achieved high classification accuracy (kappa-values 0.838-0.938) on a validation set of 180 images.
• Class activation maps identified image areas crucial for CNN decision-making.
• The algorithm successfully deciphered overlapping glomerular disease patterns within single glomeruli.

Conclusions:

• The developed deep learning model is the first to reliably recognize discrete and overlapping morphological changes in glomerular lesions.
• This AI tool, focused on conventional microscopy, shows significant promise for routine diagnostic nephropathology.
• Future work will integrate multimodal data (immunohistochemistry, electron microscopy, clinical information) for enhanced diagnostic capabilities.