Section thickness is often not what the microtome says, and AI can read it from H&E

Design

Laboratory measurement study with AI model development and independent test set

Population

154 measurement sites; 357 training and 56 test image-measurement pairs

Primary outcome

Section thickness versus microtome setting, and AI inference of thickness

Effect

More than two-thirds of sections outside ±10% of setting; best model R² 0.86, mean absolute error 0.28 µm

This laboratory study measured section thickness with confocal surface profiling at 154 sites and compared it with the microtome setting. More than two-thirds of paraffin-embedded measurements fell outside 10% of the nominal value. After deparaffinisation, thickness dropped to roughly a third of the paraffin value, and the reduction varied by tissue component, including collagen, mucin, red cell-rich areas and nuclei.

The researchers then trained neural networks on 357 matched image-measurement pairs. On an independent test set of 56 images the best model reached R² 0.86 with a mean absolute error of 0.28 µm. Staining variation influenced the estimate.

This is a methods study, not a diagnostic one. It shows that thickness is a variable, hidden preanalytical factor that may affect image-based AI and perhaps staining intensity. It does not show an effect on diagnoses.

• Do not assume that the microtome setting equals the actual section thickness.
• Log cutting parameters and re-cut when staining looks unexpectedly pale or dark.
• Ask digital pathology vendors whether their algorithms account for thickness variation.
• Treat AI thickness estimation as research for now.

Why it matters

A variable that is assumed fixed turns out to differ across a slide, and image analysis tools may be reading it as biology.

Don't overread it

The study did not test whether thickness variation changes diagnoses or AI accuracy in practice.

The statistics, in plain English

An R² of 0.86 means the model explained about 86% of the variation in measured thickness in the test images. The test set was small and from one laboratory, so performance elsewhere may be lower.

Read the rest in the app

You have read your two free briefings this month. The app carries all 27 specialties, every morning, free — and this finding is waiting in it.

Get Daily Doctor — free

Scan to keep reading on your phone. No account needed to start.