Determine whether curated clean pools reduce false positives across centers

Determine whether replacing per-slide saturation–Otsu tissue detection with entropy-based tissue detection and marker-ink curation reliably reduces false-positive rates for diffusion-based histopathology artifact detectors on external, cross-center cohorts.

Background

The study found that an entropy-based, marker-ink-curated training pool reduced clean-slide false positives by more than sixfold within the TCGA development slides. On the external SlideInspect cohort, the reduction was only about 20%, and the prespecified confirmation criterion was not met in either the saturation-based or annotator-defined tissue region.

The authors therefore leave unresolved whether the observed benefit generalizes reliably across institutions and cohorts. The discrepancy may reflect factors other than stain variation, including differences in organ composition, magnification, or other cross-center characteristics.

References

On SlideInspect (Scotto et al., 2026), a single comparison of the model trained on the curated pool (all 16 slides) against the reference model lowered the clean-control FP from 0.2264 to 0.1766 in the saturation tissue region (ratio 0.780, $97.5\%$ CI 0.505-1.002) and from 0.2305 to 0.1857 in the annotator's tissue region (ratio 0.806, 0.671-0.988); neither met our criterion for confirmation (Table 2, rows 9 and 10; Table 4).

— Tissue Detection Determines False Positives in Diffusion-Based Histopathology Artifact Detection  (2609.40083 - Moutselos et al., 30 Sep 2026) in Sections 3.7 and 4.3; Limitations, Section 4.6

One reading, which we did not test, is that the two detectors define normal at different levels.

— Tissue Detection Determines False Positives in Diffusion-Based Histopathology Artifact Detection  (2609.40083 - Moutselos et al., 30 Sep 2026) in Section 4.2