Probabilistic modeling of Diabetic Nephropathy progression

Area of Science:

• Nephrology
• Computational Pathology
• Medical Imaging Analysis

Background:

• Diabetic nephropathy (DN) staging relies on pathologist assessments of glomerular changes.
• Current methods are semi-qualitative and glomerulus-specific.
• Understanding DN progression drivers requires quantitative analysis of histological features.

Purpose of the Study:

• To develop a probabilistic model for inferring diabetic nephropathy (DN) stage from glomerular image features.
• To integrate traditional and novel image-derived features for improved DN classification.
• To investigate the influence of abstract spatial features on DN staging.

Main Methods:

• Bayesian Network (BN) construction using the structural Hill-Climbing algorithm.
• Incorporation of traditional glomerular features (e.g., area, nuclei count).
• Utilization of Minimum Spanning Trees (MST) for abstract spatial feature quantification.
• Querying the BN with Markov Particle filters.

Main Results:

• The model successfully inferred DN stage membership using a combination of features.
• Abstract features derived from MST showed variable influence on DN stage across disease progression.
• Results were validated using multi-institutional image data.

Conclusions:

• Probabilistic inference of DN stage from glomerular images is feasible.
• Abstract image features provide valuable, dynamic insights into DN progression.
• This approach offers a "white box" quantitative tool for nephrologists.