Unsupervised learning for labeling global glomerulosclerosis.

Area of Science:

• Nephropathology
• Computational Pathology
• Medical Image Analysis

Background:

• Supervised learning in nephropathology requires time-consuming manual image labeling by experts.
• Unsupervised strategies like glomeruli clustering can reduce this labeling bottleneck.
• Previous studies suggested clustering for semi-automated or unsupervised training, but its accuracy for basic glomeruli classification remained unclear.

Purpose of the Study:

• To evaluate the accuracy and limitations of clustering for separating globally sclerosed and non-globally sclerosed glomeruli.
• To address the gap in understanding clustering performance for basic glomeruli classification.

Main Methods:

• Clustering was applied to 10 diverse labeled nephropathology datasets.
• Feature embeddings from 34 different pre-trained Convolutional Neural Network (CNN) models were utilized.

Main Results:

• Clustering demonstrated high feasibility for separating globally and non-globally sclerosed glomeruli.
• Accuracies exceeding 95% were achieved in the majority of tested datasets.

Conclusions:

• Clustering is a highly accurate method for pre-labeling glomeruli, serving as a strong foundation for downstream interactive or unsupervised learning.
• These findings advance the development of clinically applicable glomerular classification models.
• Future work should explore domain-specific feature extractors via contrastive learning or foundation models for enhanced performance.