Computational Segmentation and Classification of Diabetic Glomerulosclerosis.

Area of Science:

• Nephrology
• Digital Pathology
• Machine Learning

Background:

• Diabetic nephropathy (DN) diagnosis relies on pathologist visual classification of glomerular lesions.
• Inter-pathologist variability can affect diagnostic consistency.
• Digital algorithms offer potential to standardize interpretation and reduce variability.

Purpose of the Study:

• To develop a digital pipeline for classifying renal biopsies in diabetic nephropathy.
• To combine traditional image analysis with machine learning for efficient and accurate structural quantification.
• To reduce manual effort and incorporate biological prior information into the classification model.

Main Methods:

• Developed a digital pipeline integrating image analysis and machine learning for DN classification.
• Simplified glomerular structure into nuclei, capillary lumina, Bowman spaces, and PAS-positive structures.
• Utilized convolutional neural networks for boundary and nuclei detection, and unsupervised techniques for other structures.
• Defined digital features quantifying DN progression and employed a recurrent network for classification.

Main Results:

• Digital classification showed moderate agreement with a senior pathologist (κ = 0.55).
• Agreement with two other pathologists was κ1 = 0.68 and κ2 = 0.48.
• High accuracy was achieved in detecting glomerular boundaries (0.93±0.04), nuclei (0.94 sensitivity, 0.93 specificity), and structural components (0.95 sensitivity, 0.99 specificity).

Conclusions:

• Computational approaches are comparable to human visual classification in assessing diabetic nephropathy.
• Digital classification can enhance precision in clinical decision-making workflows.
• Histologic image features derived computationally hold significant diagnostic value for augmenting clinical diagnostics.