Direct measurement and AI-based inference reveal histological section thickness as a variable physical property accessible from routine H&E images

Dataset and Code for Histological Section Thickness Estimation OverviewThis repository provides the dataset and code associated with the study: "Direct measurement and AI-based inference reveal histological section thickness as a variable physical property accessible from routine H&E images" Two independent computational approaches are included: 1. CNN-based regression models (VGG16, ResNet18, ResNet50)2. GAN-based method (pix2pix) These pipelines are independent and use different data formats, preprocessing strategies, and outputs. Repository Structure (Zenodo) README.txt CNN_code.zipCNN_data.zip GAN_code.zipGAN_data.zip Datalist_summary.xls Contents CNN_code.zipContains:- CNN source code (src/)- configuration files (configs/)- README.md- pyproject.toml (dependency specification) CNN_data.zipContains the dataset in a structure compatible with the CNN code: data/ ├── processed/ │ └── dataset.json (metadata, data split, labels) └── raw/ ├── HE_raw_data/ (JPG images, 1600×1200) └── datalist_hsv.xlsx (data required for thickness calculation) Total dataset: 413 image–measurement pairsModel-development dataset: 357 pairsIndependent test dataset: 56 pairs Important: - H&E images are NOT pre-divided into training/test sets- The split is defined programmatically using dataset.json- This ensures reproducibility of the exact data split used in the study GAN_code.zipContains:- preprocessing scripts for pix2pix training (image tiling)- postprocessing scripts for inference (tile reconstruction)- README_GAN.docx The pix2pix implementation is based on:https://github.com/affinelayer/pix2pix-tensorflow GAN_data.zipContains:- HE_GAN_training.zip- HE_GAN_test.zip- heatmap_training.zip- heatmap_test.zip Total dataset: 413 image–heatmap pairsModel-development dataset: 357 pairsIndependent test dataset: 56 pairs Includes:- Preprocessed PNG images (1024×1024)- Corresponding thickness heatmaps (8-bit grayscale images with pixel intensities linearly scaled to represent thickness values from 0 to 10 µm) Datalist_summary.xlsThis table provides the correspondence between image files, measured thickness values, dataset assignments, and model predictions used in the manuscript. It includes:- file names- data attributes- measured thickness values (mean tissue thickness [μm] within each image)- predicted values (CNN and GAN) Key Differences Between CNN and GAN CNN:- Input: JPG (1600×1200)- Output: single thickness value- Task: regression GAN:- Input: PNG (1024×1024)- Output: thickness heatmap- Task: image-to-image translation Relationship Between Datasets - CNN uses original H&E images- GAN uses preprocessed images derived from the originals These datasets are not identical and not interchangeable. Reproducibility - CNN dependencies are specified in pyproject.toml- GAN implementation is based on an external pix2pix repository- CNN data split is defined via dataset.json- GAN uses explicitly separated training/test datasets Usage Guidance Use CNN if:- you need quantitative thickness values Use GAN if:- you need spatial thickness maps License This dataset is released under the Creative Commons Attribution-NonCommercial 4.0 International (CC BY-NC 4.0). You are free to:- Share and adapt the material Under the following conditions:- Attribution is required- Non-commercial use only Commercial use is strictly prohibited. Citation If you use this dataset or code, please cite the associated publication. (To be added upon publication) Notes - CNN and GAN pipelines are independent implementations- Input formats are not interchangeable- Please follow the instructions in each code archive Contact For questions regarding this dataset, please contact the corresponding author.