Cervical Cancer Detection in Pap Smear Images

Pedro Maranhao, Tsang Ing Ren, George D. C. Cavalcanti

Abstract: Cervical cancer is a significant cause of global cancer mortality, with early detection playing a critical role in improving treatment outcomes. However, the traditional method of analysis through manual investigation of the Pap smear slides is time-consuming and prone to human error, making the development of consistent, automated diagnosis tools necessary. This study evaluates the efficacy of different advanced deep learning models in detecting and classifying cervical cells from Pap smear images. Using the CRIC’s dataset, aligned with the Bethesda system for cytopathological classification, a comparative analysis of three state-of-the-art object detection architectures, YOLOv9, GELAN, and RT-DETR, is conducted. The experimental setup of this research encompasses a variety of augmentation techniques, experimenting with both image-level and box-level augmentations and evaluating these models across an all-inclusive six-class approach. The models’ performances are assessed using the mean Average Precision (mAP). YOLOv9 obtained the highest mAP in the image-level augmented dataset. The code, and further information of the experiments can be found on GitHub https://github.com/PedroDidier/CervicalCancerDetectionPapSmear