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

Approximately 99,000 H&E-stained histopathological image patches from the GasHisSDB and GCHTID datasets, used to train and validate a multi-task deep learning model. The dataset was created by Qing-Chun Feng and last updated on 2026-04-21. The model achieved a classification F1-score of 0.938 and a segmentation Dice coefficient of 0.839 on the test set.

Use Cases

• Training multi-task deep learning models for joint classification and segmentation of tumor structures based on histopathological patches.
• Benchmarking model robustness against staining perturbations based on the reported 2.4% Dice decrease under such conditions.
• Evaluating model explainability techniques like Grad-CAM and LRP based on the expert-model agreement rate (EMAR) of 0.864.
• Assessing cross-dataset generalization for pathology AI based on the reported F1-score of 0.902 in transfer from GCHTID to GasHisSDB.

Strengths

• Approximately 99,000 H&E-stained image patches provide a substantial base for model training.
• Model performance is quantitatively validated with a classification F1-score of 0.938 and segmentation Dice of 0.839.
• Explainability was assessed with multiple methods, achieving a Grad-CAM correlation of r = 0.86 and expert-model agreement.
• Demonstrated cross-dataset transfer capability with an F1-score of 0.902 and robustness to staining perturbations.

Limitations

• The dataset is described in a 14.1 KB DOCX file, suggesting the actual image data is not directly included and must be sourced separately.
• Validation was conducted under controlled benchmark conditions using patch-level data; clinical feasibility requires further study on whole-slide images.
• Specific column-level documentation for any associated metadata is absent.

Provenance

Source

GasHisSDB and GCHTID datasets, aggregated via figshare.

Collection Method

Images were preprocessed with color normalization and quality control, used for training a Swin Transformer, DeepLabV3+, and R2U-Net model.

Freshness

Last updated 2026-04-21 04:11:12.