FASTER YOLO: AN EFFICIENT FRAMEWORK FOR CERVICAL CANCER CELL DETECTION WITH DEFORMABLE CONVOLUTIONAL ATTENTION

Cervical cancer is the fourth most common disease worldwide. The most common diagnostic method required for cervical cancer screening is the pap smear test. Making precise diagnosis, identifying and classifying cells and closely examining each slide all take a significant amount of time and work. Long stretches of visual inspection can make human mistakes more likely thereby resulting in incorrect classification of cells. An essential stage in automatic cytopathology diagnosis is the detection of nuclei in cervical cell images. In recent years, YOLO (You Only Look Once) models have been the most popular paradigm in the field of real-time object detection because of their successful balance between detection performance and processing cost. This work focuses on a number of YOLO models and various cutting-edge object detection methods that are trained on the popular SIPAKMED benchmark dataset. This dataset contains annotated labels for each image. In this paper, we provide an improved YOLO-based object detection model that achieves performance comparable to state-of-the-art YOLO models while dramatically decreasing computing complexity. The proposed model Faster YOLOv13s is built with an optimized attention aware architecture that prioritizes efficiency above detection accuracy. Experimental results show that the proposed model achieved a competitive mAP50 score of 87.00% compared to the best-performing YOLO model while significantly reducing the number of trainable parameters and taking significantly less training time. The results of this study are meant to guide future clinical applications and identify the best model for cervical cancer detection.