TCGA Gastrointestinal Cancer Histology for MSI/MSS Classification

The TCGA cohort provides 411,890 unique image patches from formalin-fixed paraffin-embedded diagnostic slides of colorectal and gastric cancer patients. These 224x224 pixel patches are preprocessed with tumor detection, color normalization, and are labeled for microsatellite stability status. The dataset is structured for machine learning, with patient-level randomization into training and test sets for classification tasks.

Use Cases

• [Classifying microsatellite instability status] based on [histological image patches from FFPE slides]
• [Training convolutional neural networks] based on [pre-split training and test sets with patient-level separation]
• [Benchmarking image normalization techniques] based on [Macenko-method normalized patches at 0.5 µm/px resolution]
• [Studying tumor morphology in gastrointestinal cancers] based on [tumor-detected patches from TCGA diagnostic slides]

Strengths

• Contains 411,890 unique, preprocessed image patches with clear provenance from the TCGA cohort.
• Includes detailed preprocessing steps like tumor detection, resizing, and color normalization, enhancing consistency for model training.
• Patient-level randomization for train/test splits (~70%/30%) helps prevent data leakage and supports robust model evaluation.

Limitations

• Metadata such as column names, exact row counts per file, last update date, and file size are not provided across sources.
• The description is truncated (e.g., 'CRC_DX_TEST_M'), indicating incomplete documentation of all dataset components.
• The dataset appears to be focused on a specific preprocessing pipeline, which may limit applicability to other analytical approaches.

Provenance

Source

The Cancer Genome Atlas (TCGA) cohort, via the GDC data portal.

Collection Method

Derived from whole slide SVS images of FFPE diagnostic slides, processed with automatic tumor detection, tiling, resizing, and color normalization.