Utilizing Deep Learning to Identify Electron-Dense Deposits in Renal Biopsy Electron Microscopy Images

Area of Science:

• Nephrology
• Digital Pathology
• Artificial Intelligence in Medicine

Background:

• Electron microscopy (EM) is vital for identifying glomerular deposits in kidney biopsies.
• Manual classification of these deposits is time-consuming and prone to inter-observer variability.
• Automating this process can improve diagnostic efficiency and consistency.

Purpose of the Study:

• To develop and evaluate a deep learning-based platform for automated classification of electron-dense deposit locations in EM images.
• To compare the performance of the deep learning model against human pathologists.

Main Methods:

• Retrospective collection of 4,303 EM images from 1,039 kidney biopsies.
• Ground truth established by expert pathologists categorizing deposits into mesangial, subepithelial, intramembranous, and subendothelial.
• Development of a ResNet18-based deep learning model for binary and multi-class classification.
• Validation against expert renal pathologists using Cohen's Kappa and accuracy metrics.

Main Results:

• The deep learning model achieved high accuracy in identifying deposit presence (AUC 0.959, accuracy 0.899).
• Classification subnets demonstrated strong performance for specific deposit locations (e.g., subepithelial AUC 0.987, intramembranous AUC 0.986).
• The model's accuracy surpassed comprehensive renal pathologists but was lower than EM specialists.

Conclusions:

• A web platform for automated electron-dense deposit location assessment in kidney biopsy EM images was successfully developed.
• The deep learning model provides an efficient and reliable tool, outperforming comprehensive renal pathologists.
• This technology has the potential to enhance the accuracy and speed of kidney disease diagnosis.