TLDR: This study evaluates advanced AI cell foundation models (2025), including CellViT++ variants and Cellpose-SAM, for accurate cell nuclei segmentation in challenging kidney pathology cases. Using a human-in-the-loop rating system on a curated dataset of difficult samples, the research found that CellViT++ [Virchow] had the highest standalone performance. Crucially, a fusion strategy combining these new models achieved significantly higher “Good” predictions and nearly eliminated “Bad” predictions, successfully resolving most challenging cases unaddressed by previous models. This demonstrates the enhanced robustness and precision of newer AI models and ensemble approaches in renal pathology.
Accurate segmentation of cell nuclei is a fundamental step in analyzing kidney pathology, crucial for tasks like cell counting and phenotype classification. However, the diverse morphology of renal tissues and variations in imaging have historically made this a significant challenge for traditional deep learning methods.
These traditional approaches often require extensive, pixel-level annotations and struggle to generalize across different tissue types, staining protocols, and imaging modalities. This has led to a growing interest in foundation models – large-scale, pre-trained AI architectures designed for broad adaptability across various biomedical tasks.
A prior study in 2024 evaluated earlier cell foundation models like Cellpose, StarDist, and CellViT on kidney pathology datasets. While these models showed promise, they consistently failed on particularly challenging image patches, such as those with low contrast, overlapping nuclei, or atypical morphologies. This raised a critical question: Can newly released cell foundation models overcome these persistent limitations?
Motivated by this, a recent study conducted a comprehensive evaluation of advanced AI cell foundation models released in 2025. These include variants of CellViT++ (HIPT, SAM, Virchow) and Cellpose-SAM, which leverage powerful large-scale pre-trained Vision Transformers like SAM, Virchow, and HIPT. These newer models are integrated into established cell instance segmentation frameworks, aiming for better generalization and segmentation accuracy, especially on the difficult samples identified in previous assessments.
The researchers used a standardized inference pipeline and incorporated a human-in-the-loop quality rating framework, similar to their prior work, to assess segmentation quality in a realistic clinical setting. This involved a curated dataset of 2,091 challenging kidney image patches, previously rated as “Medium” or “Bad” due to segmentation uncertainty.
Evaluating Individual Model Performance
The study first evaluated the standalone performance of each new foundation model. Among them, CellViT++ [Virchow] emerged as the strongest performer, with 40.3% of predictions rated as “Good” and only 0.9% as “Bad.” Cellpose-SAM and CellViT++ [SAM] also performed well, with 32.2% and 28.8% “Good” ratings, respectively, and very few “Bad” predictions. In contrast, CellViT++ [HIPT] had the lowest proportion of “Good” ratings (20.3%) and the highest “Bad” ratio (7.6%), indicating weaker robustness on this challenging dataset. While the majority of predictions for all models were rated as “Medium,” these differences highlight the significant improvements in instance segmentation quality offered by these new AI cell foundation models, largely due to large-scale pretraining.
The Power of Fusion: Enhanced Reliability
To explore whether combining outputs from multiple models could further enhance segmentation reliability, the researchers implemented a fused model strategy. This fused model achieved remarkable results: 62.2% “Good” predictions and only 0.4% “Bad” predictions. This represents a substantial reduction in segmentation failures compared to any individual model, demonstrating that the fusion strategy effectively mitigates individual model weaknesses without requiring additional retraining or supervision. The near-elimination of “Bad” predictions underscores the potential of fusion-based strategies to deliver more robust and dependable cell segmentation in complex kidney pathology settings.
Comparing with Previous Work
A rigorous comparison with the 2024 study revealed significant improvements. On the 2,091 challenging image patches, the 2025 fusion model dramatically increased “Good” predictions from 6.4% to 62.2%, while “Medium” predictions dropped from 93.5% to 37.4%. Consistent gains were also observed on the full dataset of 8,789 image patches, with “Good” predictions rising from 68.3% to 81.6%. These results confirm that fusing updated AI cell foundation models significantly improves nuclei instance segmentation precision and robustness, particularly in converting previously difficult samples into high-quality segmentation outputs.
Qualitative results further illustrated these improvements. For instance, Cellpose-SAM was shown to reduce false negatives and even detect nuclei in heavily stained regions, while also eliminating false positives from older models. The new AI foundation models demonstrated improved segmentation performance for faint nuclei and in low-contrast areas.
Also Read:
- Unlocking Insights in Kidney Cancer: A New AI Approach for Interpretable Pathology
- AttriGen: AI System Automates Detailed Blood Cell Analysis
Conclusion
This study successfully addressed a key limitation of prior foundation models in kidney cell nuclei segmentation: their failure to robustly segment hard image patches. By benchmarking state-of-the-art AI cell foundation models from 2025 and introducing an effective fusion strategy, the research demonstrated that the 2025 fusion model substantially outperforms previous models. It resolved the vast majority of challenging cases that were previously unaddressed, significantly increasing “Good” segmentation outcomes while nearly eliminating “Bad” predictions on difficult samples. These findings highlight the immense value of ensemble approaches with multiple new AI cell foundation models in enhancing generalization and robustness for organ-specific applications. The curated benchmark of hard image patches and the improved nuclei segmentation outcomes provide a practical foundation for future model refinement and deployment in renal pathology workflows. You can read the full research paper here.
