SE-DeepLabV3+: Cervical Cell Segmentation and Classification Using a Novel SE-Based DeepLabV3+ and Ensemble Method

Cervical cancer remains a major cause of morbidity and mortality worldwide, highlighting the need for early detection and effective treatment strategies. This study develops a robust deep learning-based system for cervical cell segmentation and classification using a novel Squeeze-and-Excitation DeepLabV3+ (SE-DeepLabV3+) model with a dynamic atrous rate for segmentation (instead of a fixed dilation rate) and an ensemble method for classification. The system leverages both global context and local features to enhance segmentation performance. It is important to note that for the Pomeranian and SIPaKMeD datasets, the segmentation masks were automatically generated using OpenCV without expert validation, which may limit immediate clinical applicability. The classification model utilizes an ensemble approach by combining features from multiple pre-trained convolutional neural networks (CNNs) and introducing learnable weights (dense layers applied to the global average pooling layer) to dynamically adjust the importance of each model’s feature and improve prediction accuracy. The performance of the proposed system is evaluated using the Pomeranian, SIPaKMeD, and Herlev datasets. For segmentation, the SE-based DeepLabV3+ outperforms the standard DeepLabV3+ across all datasets, achieving the highest accuracy in the Pomeranian dataset (99.09%), SIPaKMeD dataset (96.09%), and Herlev dataset (87.12%). In classification, the proposed ensemble method achieves an accuracy of 99% for the Pomeranian dataset, 94% for SIPaKMeD (multiclass), 99% for SIPaKMeD (binary), and 90% for Herlev in classification-only. These results demonstrate the effectiveness of the proposed SE-DeepLabV3+ model for segmentation and the ensemble classification method in improving overall cervical cell analysis, leading to more accurate and reliable predictions.