Assessment of Cell Nuclei AI Foundation Models in Kidney Pathology

Area of Science:

• Digital Pathology
• Computational Biology
• Medical Imaging Analysis

Background:

• Cell nuclei instance segmentation is vital for digital kidney pathology.
• Traditional methods struggle with generalizability across diverse datasets.
• Foundation models (FMs) offer a promising, generalizable approach for cell segmentation.

Purpose of the Study:

• To conduct a large-scale evaluation of SOTA cell nuclei foundation models in kidney pathology.
• To assess the performance of Cellpose, StarDist, and CellViT on a diverse dataset.
• To identify remaining challenges in automated nuclei segmentation for kidney tissues.

Main Methods:

• Created a diverse evaluation dataset of 2,542 human and rodent kidney whole slide images (WSIs).
• Included variations in tissue types, sizes, and staining methods for comprehensive testing.
• Quantitatively analyzed prediction distributions to identify performance gaps.

Main Results:

• Identified a persistent performance gap in kidney pathology nuclei segmentation across evaluated models.
• CellViT demonstrated superior performance compared to Cellpose and StarDist in this context.
• No single foundation model achieved perfect segmentation accuracy for kidney nuclei.

Conclusions:

• CellViT is a strong candidate for kidney pathology nuclei segmentation, but limitations exist.
• Further advancements are needed to fully bridge the performance gap in automated segmentation for kidney pathology.
• This study represents the largest-scale evaluation of FMs for kidney nuclei segmentation to date.