Deep Learning-Based Histopathologic Assessment of Kidney Tissue.

Area of Science:

• Digital pathology
• Artificial intelligence in medicine
• Renal histopathology

Background:

• Deep neural networks (DNNs) are advancing digital analysis of histopathologic images.
• Convolutional neural networks (CNNs) are being developed for complex image segmentation tasks.

Purpose of the Study:

• To train a CNN for multiclass segmentation of periodic acid-Schiff (PAS)-stained kidney tissue sections.
• To evaluate the network's performance on digitized kidney transplant biopsies and nephrectomy samples.

Main Methods:

• Training a CNN on 40 whole-slide images with multiclass annotations.
• Applying the network to four independent datasets for validation.
• Calculating Dice coefficients for ten tissue classes and comparing network measures with pathologist scoring (Banff classification).

Main Results:

• Achieved weighted mean Dice coefficients of 0.80 and 0.84 for PAS-stained kidney biopsies.
• Demonstrated high segmentation accuracy for "glomeruli" (Dice coefficients 0.95 and 0.94).
• Showed significant correlations between network-based measures and visually scored histologic components.

Conclusions:

• This is the first CNN for multiclass segmentation of PAS-stained kidney samples.
• The developed network shows potential for quantitative histopathology studies and routine diagnostics.