Transformer-based multiclass segmentation pipeline for basic kidney histology.

Area of Science:

• Renal pathology
• Computational pathology
• Medical image analysis

Background:

• Deep learning in renal pathology primarily targets morphology, with limited research on model versatility in severe kidney damage.
• Assessing model performance across different data distributions (modal/domain shift) is crucial for clinical applicability.

Purpose of the Study:

• To compare the modal/domain shift capabilities of Convolutional Neural Network (CNN)-based and Transformer-based deep learning models in renal pathology.
• To evaluate model performance in segmenting kidney histology, particularly in regions with severe damage, fibrosis, and inflammation.

Main Methods:

• Two splitting strategies (WSI-level and patch-level) were used to simulate multi-center data distribution.
• CNN- (UNet-ResNet18) and Transformer-based (M2F-Swin-B) models were trained and compared on these strategies.
• Models were validated on an external dataset, with sensitivity analysis for fibrosis and inflammation levels.

Main Results:

• M2F-Swin-B significantly outperformed UNet-ResNet18 in average Intersection over Union (A-IoU) and per-class IoU at both patch- and WSI-levels.
• M2F-Swin-B demonstrated superior performance in areas with higher degrees of fibrosis and inflammation, and achieved a higher IoU for arteries.
• The attention mechanism in Mask2Former (used in M2F-Swin-B) resulted in crisper, more uniform segmentation, especially with limited data.

Conclusions:

• Transformer-based models, particularly M2F-Swin-B, offer enhanced versatility and performance for renal histology segmentation compared to CNN-based models.
• The developed multi-class segmentation pipeline is robust for kidney histology analysis.
• Attention mechanisms in Transformer models are beneficial for segmentation quality, even in data-scarce scenarios.