MONAI lung detector preprint flags slice-thickness fragility

That’s a deployment-relevant warning: for this detector, acquisition protocol seems to matter more than modest dose changes, and no confidence threshold sweep will fix a bad input stack.

• –5 mm slices causing a 42% relative sensitivity drop suggests the model is brittle to z-axis resolution and small-lesion visibility.
• –Only about a 4 percentage point hit at 25-50% dose reduction implies low-dose screening may be a more manageable tradeoff than thick-slice reconstruction.
• –The 0.1-0.9 threshold sweep strengthens the claim that the effect is structural, not just a calibration artifact.
• –Because the setup uses LUNA16 weights on LIDC-IDRI data, the result reads like a sharp example of domain shift between benchmark-style training data and real deployment conditions.