Virtual staining in pathology: image quality and tissue fidelity

Professor Aydogan Ozcan discusses Pictor Labs

UCLA TDG · 2023-06-02 · Source

In this English-language UCLA TDG video from 2023, Aydogan Ozcan discusses virtual staining and Pictor Labs. This historical introduction presents a researcher and founder perspective; it is not a report of the 2025 studies below or independent clinical validation.

Research: an image transformation beyond adding color

In a 2025 study, Zhang and colleagues used a diffusion model to transform autofluorescence images of unstained human lung tissue into H&E-like images. Images of the same tissue after chemical staining served as the reference. The researchers reported 4–5-fold pixel super-resolution relative to lower-resolution inputs and methods to reduce variation between outputs. This is an image-generation research result, not evidence of diagnostic performance across all diseases.

Quality assessment: a realistic error is still an error

UCLA’s account of AQuA reports 99.8% accuracy distinguishing erroneous from error-free virtually stained kidney and lung images. This is not cancer diagnostic accuracy; it describes the study’s particular error-detection task. During assessment, the system used relationships with unstained images without requiring a correctly chemically stained reference. The institution also disclosed a patent application covering the technology. The finding does not guarantee equivalent performance in other laboratories.

Clinical use and regulatory authorization are separate assessments

The sources discussed here report research findings. We do not interpret them as a product authorization document or validation of clinical use at your institution. Product name, version, intended purpose and jurisdiction need separate examination when reviewing authorization. Related research groups publishing two papers does not establish independent external validation. This article makes no performance or approval claim for RAI products.

Editorial proposal: make the measured task visible

We recommend displaying image similarity, error detection and agreement in clinical decisions as separate headings in an evaluation report. A figure such as 99.8% should always identify the task, dataset and comparator being measured. Alongside attractive examples, show failures, rejected images and cases referred for human review. These distinctions help readers judge what the evidence actually supports.

An evaluation record for responsible AI

Our editorial proposal is to preserve the original image, label generated images clearly and record the model version, tissue type and imaging conditions. The reviewing pathologist should be able to identify transformed areas and the route to further verification. These suggestions are a framework for questioning research before implementation, not an institution-specific clinical protocol or individual treatment advice.