RIVA: An Image Dataset of Conventional Pap Smear Cytology with Multiple Independent Annotations
Paola Bianchi, Sol Anselmo, Malena Vásquez Currié, J. J. Medel, Estefanía Uelf, A. Dos Santos, Noemí Buosi, Rosana Vargas, Juliana Reves Szemere, Bruno Volcovinsky, Hugo Massaroli, Manuel Andrade, Alejandro Monastra, Emmanuel Iarussi, Viviana Siless, Luciana Bruno
Abstract
The Pap smear remains the primary screening test for cervical cancer in many low-resource regions, yet publicly available image datasets largely feature liquid-based preparations. We introduce RIVA, a high-resolution collection of 959 conventional-smear images (1024 × 1024 px) scanned at 40x magnification, sourced from 115 patients. To ensure label quality, each image was annotated by up to four independent medical professionals, with 42% of the images reviewed by all four, resulting in 26,158 annotations based on the Bethesda classification. Annotations provide coordinates of nuclei and classification labels by up to four annotators. The dataset includes 15,949 unique cells across five (pre)cancerous types (SCC, HSIL, ASCH, LSIL, ASCUS) and three non-lesion categories (NILM, ENDO, INFL). These four-expert annotations not only give RIVA a consensus-driven ground truth for robust AI training but also enable inter-annotator consistency analysis-agreement rates reach 94% for lesion vs. non-lesion and 74% across the full eight-category Bethesda scheme.