GasHisSDB and GCHTID: 99,000 H&E-Stained Gastrointestinal Cancer Histopathology Patches

Approximately 99,000 H&E-stained histopathological image patches from the GasHisSDB and GCHTID datasets, used to train a multi-task deep learning model for gastrointestinal cancer analysis. The dataset was created by Qing-Chun Feng and last updated in April 2026. Model validation achieved a classification F1-score of 0.938 and a segmentation Dice coefficient of 0.839.

Use Cases

• Training multi-task models for joint classification and segmentation of tumor and microenvironment structures based on the described deep learning framework.
• Benchmarking model robustness against staining perturbations based on the reported 2.4% Dice coefficient decrease under such tests.
• Evaluating model explainability methods like Grad-CAM and LRP based on the expert validation and reported correlation metrics.
• Assessing cross-dataset generalization for pathology AI based on the transfer learning results from GCHTID to GasHisSDB.

Strengths

• Approximately 99,000 annotated image patches provide a substantial base for model training.
• Performance metrics are reported, including a classification F1-score of 0.938 and segmentation Dice of 0.839.
• Includes explainability validation with an expert-model agreement rate (EMAR) of 0.864.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• The 14.3 KB file size suggests the dataset itself is not included; the document likely describes the research and results.
• Data is at the patch level; clinical feasibility requires further validation on whole-slide images as noted in the conclusion.

Provenance

Source

GasHisSDB and GCHTID datasets

Collection Method

Images were H&E-stained and underwent preprocessing including color normalization and quality control.

Time Range

null

Freshness

Last updated 2026-04-21 04:11:12; freshness should be verified.

Geography

null