Advances in kidney biopsy lesion assessment through dense instance segmentation.

Area of Science:

• Digital pathology and computational imaging
• Artificial intelligence in medical diagnostics
• Renal pathology and disease quantification

Background:

• Renal biopsies are crucial for kidney disease diagnosis, but manual lesion scoring by pathologists shows high inter-observer variability.
• Automating lesion classification in renal histopathology is challenging due to complex anatomical structures, class imbalance, and multi-label lesions.
• Existing computational models struggle to address these complexities efficiently and generically across diverse datasets.

Purpose of the Study:

• To develop a generalized computational solution for automated lesion classification in renal biopsies from various sources.
• To reduce inter-observer variability in kidney disease diagnosis through automated quantification of lesions.
• To create a robust model capable of handling multi-class, multi-scale objects and multi-label lesions within complex renal regions-of-interest (ROIs).

Main Methods:

• Introduction of DiffRegFormer, an end-to-end dense instance segmentation sub-network combining diffusion models, transformers, and RCNNs.
• The framework efficiently recognizes over 500 objects across three anatomical classes (glomeruli, tubuli, arteries) within ROIs.
• A two-sub-network approach: dense instance segmentation and subsequent lesion classification on identified object patches.

Main Results:

• DiffRegFormer achieved 52.1% Average Precision for detection and 46.8% for segmentation on Jones' silver-stained renal Whole Slide Images (WSIs).
• The lesion classification sub-network attained 89.2% precision and 64.6% recall on 21,889 object patches.
• The model demonstrated direct domain transferability to PAS-stained renal WSIs without requiring fine-tuning.

Conclusions:

• The developed computational approach offers a generalized and efficient solution for automated lesion analysis in renal biopsies.
• DiffRegFormer significantly improves the accuracy and consistency of lesion quantification, aiding in kidney disease diagnosis.
• The model's ability to transfer domains highlights its robustness and potential for broad clinical application in digital pathology.