Measuring many stain spectral hues over time reveals multiple sources of scanner color bias for AI

Introduction: AI is sensitive to inconsistency and pathology scanners make visibly different colored images. We investigated: if scanners faithfully reproduce stain color; how consistently similar-but-different hues behave over time; AI and foundation model (FM) bias by scanner-specific drift signatures; if vendor-agnostic International Color Consortium (ICC) profile calibration reduces AI bias to increase validity and QA.

Methods: Using a standardised color target with spectral relevance to many hues of stained tissue, a clinical scanner color performance was measured fortnightly alongside imaging a cohort of cancer slides. An ICC profile unique to the scanner at each timepoint was applied. Task-specific AI was compared to fine-tuned FM to investigate effects of scanner drift and response to hue-specific characteristics. Multi-vendor scanner colour variation established if standardisation by routine, high-fidelity ICC profiling eliminates AI color bias.

Results: Similar-but-different hue variability was seen that would be missed by just looking at one measure of a stain. Changes over time within human perception and across multiple similar hues included significant variation for all colors and their order of fidelity after a service, altering scanner characteristics. AI consistency degraded with cross-hue drift and FMs carried bias for images from specific time-points. Task-specific volumes of data may fine-tune out FM bias. Different cross-color infidelities were seen between scanner models as an exponential QA issue for AI beyond one device.

Conclusion: Scanner colors are high in variability across close stains, creating complex domain-shift for AI. Routinely measuring multiple hues of stains is essential to understand color drift potentially leading to time-point bias within FMs. Needing task-specific image volumes to overcome bias may negate benefits of 'big data' FMs. Device-specific ICC standardization is an unbiased method to remove vendor color 'lock-in' for AI.

Learning Objectives:

1. Appreciate that scanner color variation needs many measures of similar colors to interpret whole device accuracy
2. Understand and interpret color drift in scanners as a source of bias for AI that has post-validation and QA implications
3. Propose that vendor-agnostic, post-scanner, pre-analytic color calibration may independently solve bias for robust, safer AI