A Novel Graph Convolutional Network-Based Method for Cervical Cytology Image Classification

Accurate cervical cell classification is essential for reliable diagnosis in cervical cancer screening. While convolutional neural networks (CNN) have shown promise in classifying cervical cells, their constraint to local receptive fields fails to capture the latent relationships between cells. To address this limitation, this paper innovatively proposes a deep learning method that integrates relation-aware features with local features. The method has two key characteristics: (1) graphs are constructed by applying the k-nearest neighbor algorithm to local features extracted by CNN; (2) local features are fused with the relationship-aware features generated by Graph Convolutional Networks (GCN) through feature concatenation. Experiments conducted on the Spammed dataset demonstrated that the proposed model achieves classification accuracy of 97.28%, sensitivity of 97.61%, specificity of 99.33%, and an F1 score of 97.92%. This classification performance shows strong potential for clinical decision in cervical cancer diagnosis.