Unsupervised machine learning for identifying important visual features through bag-of-words using histopathology...

Area of Science:

• Nephrology
• Computational Pathology
• Medical Imaging Analysis

Background:

• Kidney disease diagnosis relies on pathologist assessment of biopsy samples, which is often subjective and lacks reproducibility.
• Existing methods for analyzing histopathology images are qualitative or semi-quantitative, limiting their predictive power for patient outcomes.
• Interobserver variability in interpreting kidney biopsy slides poses a significant challenge in diagnosing and managing chronic kidney disease (CKD).

Purpose of the Study:

• To develop an unsupervised machine learning model for objective analysis of kidney histopathology images.
• To identify novel morphological features in kidney biopsies that predict CKD presence and progression.
• To overcome the limitations of subjective visual classification and improve reproducibility in kidney disease diagnostics.

Main Methods:

• An unsupervised bag-of-words model was applied to 107,471 histopathology images from 161 CKD patient biopsy cores (C-PROBE cohort).
• The model identified and ranked morphological features predictive of CKD status and kidney function.
• Model performance was evaluated using Area Under the Curve (AUC) for predicting glomerular filtration rate (GFR) and function loss.

Main Results:

• The unsupervised model identified key morphological features highly predictive of CKD at the time of biopsy and one year later.
• The method demonstrated high reliability and reproducibility, achieving an AUC of 0.93 for predicting current GFR and future kidney function loss.
• Morphological features were ranked by importance, providing insights into diagnostic markers for CKD.

Conclusions:

• Unsupervised machine learning can effectively predict kidney function and outcomes in CKD patients using histopathology images without human input.
• The developed visual dictionary derived from machine learning captures both known and unknown clinically relevant features for disease prediction.
• This approach offers a reliable and reproducible method for objective assessment of kidney biopsy samples, potentially improving CKD management.