Self-Supervised Learning for Feature Extraction from Glomerular Images and Disease Classification with Minimal...

Area of Science:

• Digital pathology
• Artificial intelligence in medicine
• Nephrology research

Background:

• Deep learning (DL) in kidney pathology requires large labeled datasets, which are difficult to obtain.
• Limited labeled data hinders the application of DL for kidney biopsy image analysis.

Purpose of the Study:

• To evaluate the effectiveness of self-supervised learning (SSL) using DINO for analyzing kidney biopsy images.
• To assess DINO's performance in classifying kidney diseases and clinical parameters compared to traditional methods.

Main Methods:

• Applied DINO (self-distillation with no labels) to 10,423 PAS-stained glomerular images.
• Extracted and visualized glomerular features using principal component analysis (PCA).
• Performed classification tasks using k-nearest neighbor (kNN) or linear head classifiers on DINO-pretrained and ImageNet-pretrained models.

Main Results:

• DINO-pretrained models captured distinct glomerular morphological features.
• DINO-pretrained models outperformed ImageNet-pretrained models in disease classification (ROC-AUC 0.93 vs. 0.89).
• DINO demonstrated superior performance with limited labeled data and in classifying clinical parameters, confirmed by external validation.

Conclusions:

• DINO enables effective extraction of histological features from unlabeled kidney biopsy images.
• SSL with DINO significantly enhances the utility of DL for kidney disease classification.