Unveiling pathology-related predictive uncertainty of glomerular lesion recognition using prototype learning.

Area of Science:

• Nephrology
• Computational Pathology
• Artificial Intelligence in Medicine

Background:

• Glomerular lesion recognition is vital for diagnosing chronic kidney disease (CKD).
• Deep learning models struggle with lesion heterogeneity, superposition, progression, and tissue incompleteness, causing prediction uncertainty.
• Analyzing pathology-related predictive uncertainty is crucial for understanding and improving model performance.

Purpose of the Study:

• To develop and validate a novel framework for pathology-related predictive uncertainty analysis in glomerular lesion recognition.
• To investigate the relationship between predictive uncertainty and pathological properties of lesions.
• To enhance the performance of deep learning models in diagnosing CKD through uncertainty analysis.

Main Methods:

• A framework integrating prototype learning for uncertainty estimation, pathology-characterized correlation analysis, and weight-redistributed prediction rectification was proposed.
• Deep prototyping, affinity embedding, and multi-dimensional uncertainty fusion were used for uncertainty estimation.
• Expert-based and learning-based approaches were employed to characterize lesions and tissues, enabling correlation analysis.

Main Results:

• The framework facilitated efficient correlation analysis, demonstrating a strong link between predictions and pathological characteristics (c-index > 0.6, p < 0.01).
• The prediction rectification module significantly improved lesion recognition performance, with accuracy gains up to 6.36% across various metrics.
• Spearman and Pearson correlation analyses confirmed the framework's effectiveness.

Conclusions:

• The developed predictive uncertainty analysis offers a valuable method for assessing computational pathology predictions in glomerular lesion recognition.
• This approach provides a practical solution for estimating pathology-related predictive uncertainty, benefiting algorithm development and clinical applications in CKD diagnosis.