PESO: Prostate Epithelium Segmentation on H&E-Stained Whole Slide Images

62 whole-slide images of prostatectomy specimens provide a detailed view of prostate cancer at a pixel resolution of 0.48 microns. This dataset includes multiple annotation layers, such as 62 U-Net-generated training masks and 25 manually corrected masks for artifact removal. It is designed for developing and testing automated segmentation algorithms for prostate epithelium and cancerous regions.

Use Cases

• Training deep learning models for epithelium segmentation using the provided 62 training masks.
• Evaluating segmentation algorithm performance on the 160 test regions exported as PNGs.
• Removing stain artifacts from histopathology images using the subset of 25 manually corrected masks.
• Studying the correlation between immunohistochemistry (P63/CK8/18) markers and H&E morphology via the color deconvolution masks.

Strengths

• Includes 62 high-resolution whole-slide images with a precise pixel resolution of 0.48 microns per pixel.
• Provides multiple annotation types, including 62 algorithm-generated masks and a subset of 25 manually corrected masks for improved accuracy.
• Contains a structured test set with 160 distinct regions, available in both standard and padded PNG formats for border analysis.

Limitations

• Key metadata such as row count, file size, and last update date are not provided across all sources.
• The dataset is split into many separate archive files, which may complicate initial data access and management.
• The description is truncated, leaving the mapping of the 160 test set files incomplete.

Provenance

Source

Radboud University Nijmegen, associated with the paper at https://doi.org/10.1038/s41598-018-37257-4

Collection Method

Whole-slide images of prostatectomy specimens, with masks generated by a trained U-Net and subsequently manually corrected for a subset.