Explain detector-specific sensitivity to clean-pool composition

Establish why changing the tissue-detection method and clean-pool composition substantially reduces false positives for the PixCell diffusion detector but does not produce a comparably large reduction for the UNI2-h nearest-neighbour anomaly detector.

Background

The diffusion detector showed a large reduction in clean-slide false positives when its training pool was rebuilt using entropy-based tissue detection. In contrast, a nearest-neighbour detector based on UNI2-h features showed only a modest reduction that failed the prespecified criterion and failed on different slides.

The paper suggests that the two detector families may define normality at different levels—reconstruction behavior for the diffusion model versus semantic feature proximity for the nearest-neighbour model—but does not test this explanation. One particularly anomalous result is the nearest-neighbour detector's very high false-positive rate on the frozen section.