AnnoCerv: A new dataset for feature-driven and image-based automated colposcopy analysis

Abstract

Abstract Colposcopy imaging is pivotal in cervical cancer diagnosis, a major health concern for women. The computational challenge lies in accurate lesion recognition. A significant hindrance for many existing machine learning solutions is the scarcity of comprehensive training datasets. To reduce this gap, we present AnnoCerv: a comprehensive dataset tailored for feature-driven and image-based colposcopy analysis. Distinctively, AnnoCerv include detailed segmentations, expert-backed colposcopic annotations and Swede scores, and a wide image variety including acetic acid, iodine, and green-filtered captures. This rich dataset supports the training of models for classifying and segmenting low-grade squamous intraepithelial lesions, detecting high-grade lesions, aiding colposcopy-guided biopsies, and predicting Swede scores – a crucial metric for medical assessments and treatment strategies. To further assist researchers, our release includes code that demonstrates data handling and processing and exemplifies a simple feature extraction and classification technique.