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

Area of Science:

• Digital pathology
• Artificial intelligence in medicine
• Computational pathology

Background:

• Deep learning (DL) shows promise in digital kidney pathology.
• DL application is limited by the need for large, expertly labeled datasets.
• Creating labeled kidney biopsy datasets is time-consuming and requires specialized expertise.

Purpose of the Study:

• To apply self-supervised learning (SSL) using DINO (self-distillation with no labels) to kidney pathology.
• To evaluate DINO's effectiveness in extracting glomerular features and classifying kidney diseases.
• To compare DINO's performance against traditional supervised methods, particularly with limited labeled data.

Main Methods:

• Applied DINO SSL to 10,423 PAS-stained glomerular images from 384 kidney biopsies.
• Visualized DINO-extracted features using principal component analysis (PCA).
• Performed classification tasks using k-nearest neighbor or linear head classifiers on DINO- and ImageNet-pretrained models, evaluating with ROC-AUC.

Main Results:

• PCA confirmed DINO's ability to capture distinct glomerular morphologic features.
• DINO-pretrained models achieved higher diagnostic accuracy (ROC-AUC=0.93) than ImageNet-pretrained models (ROC-AUC=0.89).
• DINO maintained superior performance with limited data (ROC-AUC=0.88) compared to ImageNet models (ROC-AUC=0.76), validated on external datasets.

Conclusions:

• DINO successfully extracts relevant histologic features from unlabeled kidney biopsy images.
• SSL with DINO enables effective disease classification in digital pathology, overcoming data limitations.
• DINO pretraining offers a robust approach for kidney disease analysis, especially in resource-constrained settings.