Cervical whole-slide images dataset for multiclass classification

Background: The clinical pathway for the prevention and treatment of cervical cancer depends on cytology and then the assessment of biopsy specimens, fragments of tissue removed for histological examination. This can be a significant workload and is an obvious exemplar to explore triage based on machine learning analysis of slides. Limited access to large annotated datasets of human diseased tissue is a major obstacle to developing standards and algorithms that can assist diagnosis.

Results: We present a dataset comprising 2,539 whole-slide images of cervical biopsy specimens, each annotated by several pathologists and consensus on diagnosis and individual features agreed. Each whole-slide image represents 1 slide per patient in iSyntax format, with manual annotations by pathologists in Jason format. Each whole-slide image is assigned a category label, which is the final diagnosis of the image, and a subcategory label, which declares in which subcategory the image is found.

Conclusion: This dataset has been used to build a model that accurately predicts diagnosis, allowing the possibility of automatically triaging biopsy specimens, so that the most significant pathologies can be identified rapidly and those patients selected for immediate treatment. The level of annotation, at the subslide level, and the number of cases are unique in public databases and should allow investigators to explore multiple aspects of computer vision relevant to human tissue diagnosis, with no limitation placed on access to the whole-slide images.