Direct Measurement and AI-based Inference Reveal Histological Section Thickness as a Variable Physical Property Accessible from Routine H&E Images.

PURPOSE: Variability in routine hematoxylin and eosin (H&E) images is an emerging concern in diagnostic and computational pathology. Although staining variability has been extensively studied, histological section thickness remains a largely unmeasured physical property that may influence image appearance. We investigated how section thickness varies within and between histological sections and whether it can be inferred from H&E images using artificial intelligence (AI). MATERIALS AND METHODS: We integrated high-resolution confocal surface profiling with matched H&E imaging. At 154 measurement sites, section thickness was compared with microtome preset values and between the paraffin-embedded and deparaffinized states. AI models were developed using 357 matched image-measurement pairs and evaluated in an independent test set of 56 images. RESULTS: Paraffin-embedded section thickness frequently deviated from microtome preset values, with more than two-thirds of measurements falling outside ±10% of the nominal setting. After deparaffinization, section thickness decreased to approximately one-third of the paraffin-embedded thickness. Spatial thickness maps further revealed tissue component-dependent thickness reduction, including in collagen, mucin, erythrocyte-rich areas, and nuclear structures, contributing to marked spatial heterogeneity in deparaffinized section thickness. Among the convolutional neural network-based regression models, the best-performing ResNet50 achieved a coefficient of determination of 0.86 and a mean absolute error of 0.28 μm in the independent test set. A generative adversarial network further recapitulated spatial patterns of thickness variation from H&E images. Digital color-perturbation analyses showed that staining-related image variation can influence thickness estimation. CONCLUSIONS: These findings demonstrate that histological section thickness is not merely a microtome setting but a variable, tissue-dependent, and AI-inferable physical property of routine H&E sections. Section thickness may therefore represent an underrecognized preanalytical source of image variation and a potential target for AI-assisted, thickness-aware quality control in digital pathology.