Detection of cervical cancer cells in complex situation based on improved YOLOv3 network

Cervical cancer is one of the major diseases that seriously threaten women’s health. Cervical cancer automatic screening technology is of great significance to reduce the incidence of cervical cancer. However, the current method has shortcomings: low efficiency, low accuracy, and weak generalization ability, especially in complex situation. This paper innovatively applies the YOLO algorithm to the detection of abnormal cervical cells to ensure the rapidity and accuracy of the detection. For cellular classification of small targets, complex background and irregular shapes, we add the dense block and S3Pool algorithm on the basis of the feature extraction network Darknet-53 to improve the generalization ability of the model to cell features. The improved algorithm k-means++ is used to replace the clustering algorithm k-means in the original yolov3 to cluster the target frame of the cell data set, set reasonable anchors size, reconstruct the prediction scale creatively. Moreover, the Focal Loss and balanced cross entropy function are employed to improve the detection effect of the model against complex backgrounds, tight cell clusters, and uneven number of cell types. The NMS algorithm with linear attenuation is used to post-processing the model to improve the detection accuracy of cells in the occlusion situation. Experimental verification shows that the network achieved MAP of 78.87%, which is 8.02%, 8.22% and 4.83% higher than SSD (Single Shot Multi-Box Detector), YOLOv3(You Only Look Once) and ResNet50. The optimization method proposed in this paper improves the network sensitivity and the overall accuracy, especially in complex background. The research in this paper will have significance for the future design of an automatic cervical cancer diagnosis system.