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

This dataset accompanies the paper "Comprehensive evaluation of cross-cancer generalization in histopathology segmentation models across 21 tumor types."

It contains the evaluation data used to assess the cross-cancer generalization performance of five deep learning segmentation models (trained on breast, colon, lung, kidney, and prostate tissue). Segmentation quality was evaluated by a board-certified pathologist on regions of interest (ROIs) extracted from whole-slide images across 21 TCGA tumor types. The dataset includes pathologist quality ratings, inter-rater reliability data, quantitative segmentation accuracy metrics, and clinical metadata.

Contents

scoring_data.csv — Primary rater's quality scores (0–10) for 7,616 tumor and 3,296 normal tissue ROIs across all five models and 21 TCGA projects (54,560 rows).

scoring_data_second_rater.csv — Second rater's scores for a subset of 598 segmentation outputs across all five models, used for inter-rater reliability analysis (ICC and Spearman's rank correlation).

pathologist_inference_ratings.csv — Manual quality scores on a separate validation cohort of 346 ROIs from the four model training cohorts (breast, colorectal, lung, and prostate), where ground truth annotations were available. Used for comparison against automated Dice coefficients.

dice-results/ — Dice coefficients computed on the same validation cohorts by comparing model segmentation masks against ground truth annotations exported from QuPath (4 cohorts × 4 models, 16 files).

clinical-data/ — TCGA clinical metadata for 22 cancer types (TSV format, sourced from the GDC Data Portal).

Related datasets

Code (annotation, inference, scoring, and analysis pipeline): 10.5281/zenodo.18520078

Tissue ROIs extracted from TCGA whole-slide images: 10.5281/zenodo.18668580

Segmentation masks produced by the five models: 10.5281/zenodo.18669667