Cross-Species Data Integration for Enhanced Layer Segmentation in Kidney Pathology.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Nephrology

Background:

• Accurate kidney layer segmentation is vital for diagnosing renal diseases.
• Deep learning models require extensive annotated data, which is difficult to obtain due to privacy and scarcity.
• External datasets can introduce noise, hindering model generalization.

Purpose of the Study:

• To investigate the effectiveness of using cross-species homologous data (mouse kidneys) to improve deep learning models for human kidney layer segmentation.
• To address the challenge of limited annotated clinical data for training AI models.

Main Methods:

• Jointly trained Convolutional Neural Network (CNN) and Transformer-based semantic segmentation models using both human and mouse kidney datasets.
• Utilized Periodic Acid-Schiff (PAS) stained mouse kidney data, chosen for its structural and feature similarity to human kidneys.

Main Results:

• Incorporating mouse kidney data led to an average increase in mean Intersection over Union (mIoU) of 1.77% for the renal cortex and 1.24% for the medulla.
• Dice scores improved by 1.76% for the cortex and 0.89% for the medulla.
• The approach enhanced the generalization ability of the segmentation models.

Conclusions:

• Cross-species homologous data serves as a valuable, low-noise training resource for improving AI model performance in kidney layer segmentation.
• This method is particularly effective in scenarios with limited clinical samples, offering a practical solution for enhancing diagnostic tools.