Yolov11 de baixa latência para citologia cervical: viabilizando detecção em tempo real através de microscopia apoiada por computação em borda

Abstract

Cervical cancer is the fourth most common malignant neoplasm among women worldwide and is predominantly associated with infection by the human papillomavirus (HPV). Given its high incidence, early diagnosis is essential, as symptoms are often silent or minimally perceptible. Although the conventional Papanicolaou test remains widely used, particularly in less developed regions, it presents important limitations, including a notable susceptibility to false-negative results. To improve diagnostic accuracy and efficiency, this study proposes the development of an automated system for cervical cytology analysis, integrating a CMOS microscope camera with artificial intelligence–based image detection algorithms. For this purpose, the nano (n) and small (s) variants of the YOLOv11 model were employed to detect cervical cells under real-world operating conditions. Model training was performed using publicly available datasets, including SIPaKMeD, CRIC Cervix, and RepoMedUNM. The YOLOv11n model demonstrated robust performance, achieving a mean Average Precision (mAP) of 99.30%, a recall of 99.20%, and a processing rate of 21.36 frames per second. Furthermore, comparative experiments were conducted using two datasets containing identical images with resolutions of 1280 × 1280 and 2016 × 2016 pixels. The objective was to evaluate the impact of employing resolutions higher than 640 × 640 on the performance of the (n) and (s) variants, enabling improved efficiency through the analysis of higher-definition images.