PatchCamelyon: 327,680 Histopathology Images for Metastasis Detection

327,680 color images extracted from histopathologic scans of lymph node sections form a benchmark for machine learning. Each 96x96 pixel image is annotated with a binary label indicating the presence of metastatic tissue. The dataset, created by 1aurent and last updated on Hugging Face in May 2024, is designed to be trainable on a single GPU.

Use Cases

• Binary classification of metastatic tissue based on histopathology image patches.
• Benchmarking model performance on a dataset larger than CIFAR10 but smaller than ImageNet.
• Training lightweight computer vision models suitable for single-GPU environments.

Strengths

• Contains 327,680 color images, providing a substantial volume of training data.
• Images are consistently sized at 96x96 pixels, simplifying model input pipelines.
• Explicitly designed as a benchmark for machine learning, suggesting a clear evaluation purpose.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Data may reflect bias inherent to the source histopathologic scans, such as from specific patient populations or scanning equipment.
• Description metadata is limited; actual data quality requires manual inspection after download.

Provenance

Source

Histopathologic scans of lymph node sections.

Collection Method

Extracted image patches.

Freshness

Last updated 2024-05-25 17:12:33; freshness should be verified.