How Safe AI Works in Pathology

At Pictor Labs, our AI systems are designed to augment productivity for pathologists. Virtual stains are reproducible, reviewable, and developed with human oversight at every stage. Pathologists remain central to interpretation, validation, and decision-making.

Our research-use-only virtual staining technology is built on a rigorous performance evaluation framework aligned with established digital pathology standards. Studies have demonstrated high concordance between virtual and traditional chemical stains, while preserving the interpretive role of qualified pathologists.

Virtual staining enables targeted examination of specific regions of interest while keeping the original tissue intact. Every AI-generated image is designed to be transparent, measurable, and subject to expert review. These evaluation methods ensure that virtual stains generated by AI remain reliable, reviewable, and aligned with established pathology standards.

Performance Evaluation Framework

Quantitative Metrics

Image-Based Performance

Model development and evaluation incorporate objective image-based metrics, such as Peak Signal-to-Noise Ratio (PSNR), Structural Similarity Index (SSIM), and feature-based quantitative analyses to assess image clarity, structural preservation, and color fidelity relative to reference staining.

While these metrics support model development and technical verification for certain stains, standard release criteria rely on comprehensive pathologist review to ensure visual and interpretive adequacy.

Quality Assessment

Human Pathologist Review

Quantitative metrics alone are not sufficient. Virtual stains must also meet the same qualitative standards applied to histochemically stained slides.

In blinded studies, board-certified pathologists independently evaluate virtual stains using established scoring systems. Assessments include:

  • Nuclear clarity (hematoxylin-equivalent quality)
  • Structural and cytoplasmic detail (eosin-equivalent quality)
  • Overall staining effectiveness
  • Image sharpness and interpretability

These evaluations reinforce that expert human review remains central to performance evaluation. AI outputs are not considered in isolation, they are assessed through the lens of clinical expertise.

Research Use Only

Diagnostic Concordance

In our research setting, diagnostic metrics are used strictly for evaluation and concordance analysis, not for autonomous clinical diagnosis.

Pathologists assess key disease indicators across virtual and traditional stains, including:

  • Nodal effacement
  • Diagnostic category
  • Sclerosis and necrosis
  • Final adjudicated differential diagnoses (DDx)
  • IHC recommendations

These studies evaluate alignment between virtual and chemical stains while maintaining pathologist-led interpretation. The AI is adjunctive in nature and does not generate diagnoses independently; it provides a stain representation that supports expert review.

Designed for Responsible Integration

Rigorous performance evaluation is essential as we explore the future clinical potential of virtual staining. Our approach prioritizes accuracy, consistency, transparency, repeatability, and reproducibility to meet high pathology standards.

By combining quantitative performance metrics with expert human evaluation, we ensure that AI-generated stains are reliable, reviewable, and aligned with established digital pathology standards, while keeping pathologists firmly in control of interpretation.

Virtual staining expands what is possible in digital pathology. It enhances workflow flexibility and standardization, while preserving the central role of clinical expertise in every decision.