Digital Pathology Data Set of Skin Cancers Represents Countless Hours of Hand Annotations

By: Celeste L. Dixon
Posted: Friday, February 11, 2022

To advance digital pathology machine learning capabilities to correctly identify non-melanoma skin cancers—and potentially for other research purposes—an Australian team has created and made available 290 hand-annotated histopathology tissue sections of the three most common skin cancers: basal cell carcinoma, squamous cell carcinoma, and intraepidermal carcinoma. According to Nicholas A. Hamilton, PhD, of the Institute for Molecular Bioscience at the University of Queensland, and colleagues, with the data set, researchers can compare and benchmark results when performing different skin cancer image analysis tasks, including image segmentation, classification, margin detection and measurement, specimen orientation, and assessment of cancer invasiveness. More on their data set can be found in an article published in Data in Brief.

The data set took more than 250 hours of manual annotation and curation to complete, explained the team, and it represents typical cases of the three cancer types across shave, punch, and excision biopsy contexts. “High-quality hand annotations are costly to produce and rare in current digital pathology repositories,” Dr. Hamilton and colleagues stated. The data set was produced using samples alone with clear, unambiguous diagnostic features, and the cases represented patients with a median age of 70 years, two-thirds of whom were men.

The images are identified not only by carcinoma types. “Each image is accompanied [by] a segmentation mask [that] characterizes the section [by] tissue type…including keratin, epidermis, papillary dermis, reticular dermis, hypodermis, inflammation, glands, and hair follicles,” detailed the researchers.

Such densely labeled segmentation data are necessary for the best possible training of effective machine-learning models, according to the team. What’s more, the segmentations could also be used for purposes beyond machine learning. Dr Hamilton and colleagues suggested the segmentations “may provide an excellent starting point for further annotation if researchers wish to include more tissue (sub) classes.”

Disclosure: The study authors reported no conflicts of interest.

Data in Brief