Code for: Comprehensive evaluation of cross cancer generalization in histopathology segmentation models across 21 tumor types

This repository contains the source code accompanying the paper "Comprehensive evaluation of cross-cancer generalization in histopathology segmentation models across 21 tumor types." The code implements a complete end-to-end pipeline for evaluating the cross-cancer generalization performance of five organ-specific deep learning segmentation models (breast, colon, lung, kidney, prostate) on 7,706 whole-slide images from 21 TCGA tumor types. It covers all stages from tissue annotation and model inference to manual quality scoring and statistical analysis. Components qupath-annotation/ — Groovy scripts for QuPath (v0.5.0) to create rectangular tumor/normal annotations on whole-slide images and export tissue regions. segmentation-pipeline/ — Python-based patch-wise semantic segmentation inference using a UNet architecture with EfficientNet-B0 encoder (segmentation_models_pytorch), plus Dice coefficient computation against ground-truth labels across four validation cohorts (BRCA, COADREAD, LUNG, PRAD). scoring-app/ — FastAPI web application for manual side-by-side quality assessment of five model outputs per tissue region, with zoomable viewers (OpenSeadragon), keyboard-driven scoring on a 0–10 scale, and CSV export of results. scoring-analysis/ — R/Quarto notebooks generating all publication figures and statistical analyses, including inter-rater reliability (ICC), Dice score distributions, and cross-cancer performance comparisons. Related datasets Evaluation data (scoring results, Dice coefficients, clinical metadata): 10.5281/zenodo.18518811 Tissue ROIs (exported tumor and benign ROIs from 21 TCGA cohorts): 10.5281/zenodo.18668580 Segmentation masks (model prediction masks for all ROIs): 10.5281/zenodo.18669667