Evaluation Kidney Layer Segmentation on Whole Slide Imaging using Convolutional Neural Networks and Transformers.

Area of Science:

• Digital pathology
• Renal pathology
• Medical image analysis

Background:

• Manual segmentation of kidney layer structures in whole slide images (WSI) is labor-intensive and impractical for large-scale digital pathology.
• Deep learning methods are emerging in digital renal pathology, but few have been applied to kidney layer structure segmentation.

Purpose of the Study:

• To assess the feasibility of deep learning approaches for kidney layer structure segmentation.
• To compare the performance of Convolutional Neural Network (CNN) and Transformer segmentation models on this task.

Main Methods:

• Employed representative CNN (U-Net, PSPNet, DeepLabv3+) and Transformer (Swin-Unet, Medical-Transformer, TransUNet) segmentation models.
• Quantitatively evaluated six deep learning models on renal cortex layer segmentation using mice kidney WSIs.
• Utilized Mean Intersection over Union (mIoU) as the primary evaluation metric.

Main Results:

• Transformer-based models generally outperformed CNN-based models in kidney layer segmentation.
• The evaluated deep learning models achieved a decent Mean Intersection over Union (mIoU) index.
• Demonstrated compelling advancements in automated segmentation of renal cortical structures.

Conclusions:

• Deep learning approaches are feasible and effective for kidney layer structure segmentation.
• Transformer models show superior performance compared to CNN models for this specific task.
• Automated segmentation holds promise for empowering medical professionals with more informed insights in renal pathology.