SPP-SegNet and SE-DenseNet201: A Dual-Model Approach for Cervical Cell Segmentation and Classification

Abstract

Background/Objectives: Cervical cancer, the fourth most common malignancy in women worldwide, continues to pose a significant threat to global health. Manual examination of the Pap smear image is time-consuming, labor-intensive, and prone to human error due to the large number of slides and subjective judgment. This study proposes a novel SegNet-based spatial pyramid pooling (SPP-SegNet) deep learning model for segmentation and a Squeeze-and-Excitation-based (SE-DenseNet201) model for classification, aimed at improving the accuracy of cervical cancer detection. Methods: The model incorporates the SPP bottleneck and atrous convolution in the SegNet framework, allowing for the extraction of multiscale spatial features and improving segmentation performance. The segmentation output is used as input for the classification task. The proposed method is evaluated on the Pomeranian and SIPaKMeD datasets. Results: Segmentation results show that SPP-SegNet achieves 98.53% accuracy on the Pomeranian data set, exceeding standard SegNet, 97.86%. It also achieves 94.15% accuracy on the SIPaKMeD dataset, outperforming the standard SegNet, which is 90.95%. For classification, SE-DenseNet201 achieves 93% and 99% accuracy for the Pomeranian and SIPaKMeD binary classification, respectively, using the bounding box input. Conclusions: These results show that SPP-SegNet and SE-DenseNet201 can potentially automate cervical cell segmentation and classification, facilitating the early detection and diagnosis of cervical cancer.