Kidney Week

Abstract: SA-PO0047

Evaluating Zero-Shot Claude Vision for Kidney Biopsy Adequacy Assessment

Session Information

• Artificial Intelligence: Kidney Imaging, Pathology, and Precision Phenotyping
  October 24, 2026 | Location: Exhibit Hall A, Convention Center
  Abstract Time: 10:00 AM - 12:00 PM

Category: Artificial Intelligence, Digital Health, and Data Science

• 300 Artificial Intelligence, Digital Health, and Data Science

Authors

• Eigbire-Molen, Odianosen J., Arkana, Little Rock, Arkansas, United States

Odianosen J. Eigbire-Molen, MD

• Dercatshoff Pereira Rodrigues, Ricardo, Arkana, Little Rock, Arkansas, United States

Ricardo Dercatshoff Pereira Rodrigues

• Cassol, Clarissa Araujo, Arkana, Little Rock, Arkansas, United States

Clarissa Araujo Cassol, MD

• Sharma, Shree G., Arkana, Little Rock, Arkansas, United States

Shree G. Sharma, MD, FASN

Background

Kidney biopsy inadequacy has increased from approximately 2% in 2005 to 14% in 2020, with important consequences for deceased-donor organ utilization. Previous smartphone-based adequacy assessment tools required custom-trained deep learning pipelines, including U-Net segmentation and classification (Eigbire-Molen et al., J Pathol Inform 2024). We hypothesized that Claude Vision (Anthropic), a general-purpose multimodal AI, could achieve comparable performance using zero-shot image classification guided by domain-specific prompting — without model training, GPU infrastructure, or segmentation workflow.

Methods

We evaluated 747 iPhone macro images of deceased-donor kidney biopsy cores. Each image was labeled for percent cortex by at least two nephropathologist based on PAS-stained sections, with adequacy defined as at least 30% renal cortex. Images were divided into a 600-image prompt development set and a 147-image blinded test set.

The prompt development set was used to refine a text prompt describing visual criteria for renal cortex, medulla, and perirenal fat. The blinded test set was enriched for inadequate biopsies (40 inadequate, 107 adequate) to ensure sufficient representation of the clinically higher-stakes error class. Images were classified using Claude Opus-4-7 via the Anthropic API.

Results

On the test set, Claude Vision achieved 86% accuracy, 97% sensitivity for inadequate cores, 85% specificity, and AUC-ROC of 0.91. These results compare favorably with previously reported performance from a task-trained deep learning pipeline which achieved 81% accuracy, 71% sensitivity, 83% specificity, and AUC-ROC of 0.80. The largest improvement was observed in sensitivity for inadequate biopsy cores, the clinically higher-stakes error category, with implications for reducing the rate of missed inadequate cores that may lead to erroneous organ utilization decisions.

Conclusion

Claude Vision exceeded the previously reported performance of a custom-trained deep learning pipeline for assessing deceased-donor renal biopsy adequacy from smartphone macro images. This approach can be deployed through a single API call, offering a low-barrier alternative for laboratories without machine learning expertise. Limitations include evaluation limited to deceased-donor kidneys, AI optimism bias, reduced interpretability compared with task-specific models, and susceptibility to performance changes with future model updates.