YOLO-TCT: An Effective Network For Long-Tailed Cervical Cell Detection

Abstract

The Thinprep Cytologic Test (TCT) is a vital component in the early detection of cervical cancer. However, conventional manual screening methods are hindered by inefficiencies and high levels of subjectivity. This study presents YOLO-TCT, an enhanced YOLOv9 network designed for the automated detection of abnormal cervical cells in TCT smears. Our model features a novel detection head with an adaptive weighting mechanism and Hard Polarized Self Attention (HPSA) to improve the differentiation of subtle cell characteristics. Additionally, we enhance the network’s ability to distinguish between foreground and background by incorporating a detection box quality factor into the classification loss function, and we introduce class parameters into the quality focal loss to address the class imbalance. Experimental evaluations on the CQTCT dataset reveal that YOLO-TCT achieves a mean Average Precision at IoU=0.5 (AP50) of 26.37% and a mean Average Recall (mAR) of 46.10%, with a significant improvement in recall rates for tail classes by over 26%. Notably, our model operates approximately five times faster than traditional R-CNN models while achieving superior precision and recall rates. These results underscore the effectiveness of our approach for real-time cervical cell detection in the context of long-tail distribution. The source code is accessible at https://github.com/threedteam/YOLO-TCT.