Automatic screening system for conventional Pap smear analysis

This paper presents a comprehensive automatic screening system designed to assist cytological experts in the evaluation of Pap smears for cervical cancer detection. This work presents an end-to-end automatic screening system developed specifically for conventional Pap smears. The goal of the system is to reduce the burden on human experts by providing a reliable, resource-efficient, and accurate solution that can distinguish between normal and diseased smears. The proposed automatic screening system processes whole-slide images by first dividing them into 2000 × 2000 pixel tiles, then applies a YOLO-based localization algorithm to extract individual cells. These cells are subsequently classified by an ensemble of nine deep neural networks, including convolutional and transformer-based architectures, trained on the publicly available APACC dataset. The final stage involves aggregating cell-level predictions across models to generate smear-level features, which are used to train a LightGBM classifier. The system is evaluated on a private dataset of 339 expert-annotated smears collected in a clinical setting. Despite the relatively small number of smears, over 12 million individual cell predictions were utilized, resulting in an average smear-level classification accuracy of 88.8% using cross-validation. These findings demonstrate the feasibility and potential of such automated systems to support existing cervical cancer screening workflows and to enhance diagnostic consistency and throughput.