TCGA-CRC-DX Histology Tiles, Color-Normalized and Split by MSI Status

Histological images of colorectal cancer tumor tissue, manually outlined and cut into 512 px tiles at an effective magnification of 0.5 µm/px. All tiles were color-normalized using the Macenko method and labeled with the parent patient's MSI status, with patients split into a 2:1 training-test ratio and training tiles undersampled for class balance. The dataset was derived from the TCGA database by Jakob Nikolas Kather of RWTH Aachen University.

Use Cases

• Train image classifiers to predict MSI status based on histological tile appearance.
• Benchmark color normalization techniques for histopathology images using the Macenko method.
• Develop models for tumor region segmentation from manually outlined tissue samples.
• Evaluate the impact of class balancing via random undersampling on model performance.
• Study the visual patterns distinguishing MSI-High from MSI-Low/MSS colorectal cancer cases.

Strengths

• Images are color-normalized with the Macenko method, which likely reduces staining variation.
• Patients were split into training and test sets in a defined 2:1 ratio.
• Training tiles were randomly undersampled to equalize classes, which can mitigate class imbalance.
• Tiles are derived from manually outlined tumor regions, focusing analysis on relevant tissue.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Row count is unknown, which may limit suitability assessment.
• The test set was not undersampled, which may affect performance metrics compared to the balanced training set.

Provenance

Source

The Cancer Genome Atlas (TCGA) database

Collection Method

Tumor tissue was manually outlined, the region was cut into tiles of 256 µm edge length, saved as 512 px images, and color-normalized.