Cervical cell classification with graph convolutional network.

Area of Science:

• Medical imaging analysis
• Computational pathology
• Machine learning for healthcare

Background:

• Cervical cell classification is crucial for early cervical cancer screening.
• Conventional methods rely on hand-crafted features, while Convolutional Neural Networks (CNNs) use learned deep features.
• CNNs may overlook latent image correlations, potentially limiting feature representation ability.

Purpose of the Study:

• To propose a novel cervical cell classification method using Graph Convolutional Networks (GCNs).
• To explore potential relationships among cervical cell images to enhance classification performance.
• To improve the representation ability of CNN features by incorporating relation-aware features.

Main Methods:

• Clustering CNN features of cervical cell images to reveal intrinsic relationships.
• Constructing a graph structure to capture correlations among clusters.
• Applying GCN to propagate node dependencies and generate relation-aware features.
• Integrating GCN features with CNN features to improve discriminative ability.

Main Results:

• The proposed GCN-based method demonstrates feasibility and effectiveness on the SIPaKMeD dataset.
• Experiments on a large-scale Motic liquid-based cytology dataset show superior performance under consistent and different staining conditions.
• The method outperforms existing state-of-the-art approaches in quantitative metrics like accuracy, sensitivity, specificity, and F-measure.

Conclusions:

• Exploring intrinsic relationships of cervical cells significantly improves classification accuracy.
• Relation-aware features generated by GCN effectively enhance CNN feature representation.
• The proposed method achieves superior classification performance and shows potential for automated cervical cytology screening systems.