COBRA data toolkit: A whole-slide image dataset of basal cell carcinoma and diverse skin malignancies for computational dermatopathology

COBRA: A whole-slide image dataset of basal cell carcinoma and diverse skin malignancies for computational dermatopathology Description We introduce COBRA (Classification Of Basal cell carcinoma, Risky skin tumors and Abnormalities), a comprehensive dermatopathology dataset for computational pathology research. COBRA addresses the need for large-scale, high-quality datasets to train and validate AI models for skin cancer detection, report generation, and out-of-distribution detection. The dataset contains whole-slide images (WSIs) from Radboud University Medical Center, spanning multiple diagnostic categories including: Basal Cell Carcinoma Benign & Pre-malignant lesions Squamous Cell Carcinoma Lymphoma Melanoma Merkel Cell Carcinoma Cutaneous Metastases Rare Adnexal Tumors Key Features Multi-level annotations: Slide-level diagnostic labels for all images, pixel-level annotations for resection specimens Pathology reports: Dutch-language pathology conclusions enabling vision-language model research Pre-computed embeddings: PRISM foundation model embeddings (slide-level and tile-level) for immediate use Standardized format: All images bin-packed and resampled to 0.5 µm/pixel in pyramidal TIFF format Patient metadata: Anonymized demographic information (age, gender) Getting Started All experiment runners include a small subset of the data to verify that everything is working correctly. To access the full dataset, please refer to the documentation at the repository linked below. Repository Structure cobra/ ├── bcc-biopsies/ # BCC and benign/pre-malignant biopsies │ ├── images/ # Whole-slide images (.tif) │ ├── annotations/ # Slide-level labels │ ├── metadata/ # Patient metadata │ ├── reports/ # Pathology conclusions │ ├── postprocessing/ # PRISM embeddings │ └── thumbnails/ # Low-resolution previews ├── diverse-biopsies/ # Diverse non-BCC entities + older BCC biopsies │ ├── images/ │ ├── annotations/ │ ├── metadata/ │ ├── postprocessing/ │ └── thumbnails/ └── bcc-resections/ # BCC resections with pixel-level annotations ├── images/ ├── annotations/ # GeoJSON polygons and TIFF masks │ ├── tumor/ │ └── epidermis/ ├── metadata/ ├── postprocessing/ └── thumbnails/ Annotation Details Slide-level Labels CSV files with columns: filename, label Pixel-level Annotations (Resections) GeoJSON polygons: Vector annotations for tumor and epidermis regions TIFF masks: Single-band masks with class indices (0=background, 1=tissue, 2=tumor, 3=epidermis) Pathology Reports CSV files with columns: filename, conclusion (Dutch text) Technical Specifications Property Value Image format Pyramidal TIFF (JPEG compression) Resolution 0.5 µm/pixel Tile size 256×256 pixels Color space RGB, 8-bit per channel Intended Use COBRA supports multiple computational pathology tasks: Binary and multi-class classification of skin lesions Semantic segmentation of tissue compartments Out-of-distribution detection for clinical safety Vision-language model development with pathology reports Zero-shot classification using foundation model embeddings License Dataset: CC BY 4.0 Software: Apache License 2.0 Acknowledgments The COBRA dataset was made possible through kind donations from: Mr H.J.M. Roels donation to Oncode Institute Lionsclub Doetinchem - de Graafschap Contact Daan Geijs - Dataset creator Geert Litjens - Corresponding author ([e-mail]) Department of Pathology, Radboud University Medical Center, Nijmegen, The Netherlands diagnijmegen.nlcomputationalpathologygroup.eu