Integrating domain-specific convolutional features with Swin Transformers for transparent cervical cancer diagnostics

Vollständiger Abstract

Worum geht es in dieser Arbeit?

Introduction Cervical cancer screening relies heavily on manual cytopathological evaluation, which is subjective and prone to inter-observer variability. While deep learning offers automated alternatives, existing models often struggle to capture both global cellular context and fine-grained biological nuances. To address this, we propose CAMF-Swin, a novel hybrid architecture for precise cervical cell classification. Methods The CAMF-Swin framework integrates a hierarchical Swin Transformer backbone with four domain-specific convolutional modules to explicitly extract morphological, textural, nuclear, and cytoplasmic features. A Cross-Modal Attention Fusion (CMAF) mechanism is utilized to dynamically synthesize these diverse representations. The model was trained using stratified five-fold cross-validation and an ensemble strategy. Performance was validated on the multi-class SipakMed dataset and evaluated for generalizability on an independent, highly imbalanced Mendeley Liquid-Based Cytology (LBC) dataset. Additionally, Local Interpretable Model-agnostic Explanations (LIME) were integrated for post-hoc visual interpretability. Results The CAMF-Swin ensemble achieved a classification accuracy of 97.78% and a Matthews Correlation Coefficient (MCC) of 0.9722 on the SipakMed dataset. On the highly imbalanced Mendeley LBC dataset, the model maintained robust generalization, achieving 98.96% accuracy and an F1-Score of 0.9894. Furthermore, LIME visualizations confirmed the model’s reliance on clinically relevant biomarkers, successfully identifying features such as perinuclear halos in koilocytotic cells. Discussion These results demonstrate significant performance superiority over standard Vision Transformer variants and traditional convolutional networks, highlighting the architecture’s resilience to real-world data imbalance. By providing transparent visual evidence of its decision boundaries, the model successfully bridges the gap between neural network predictions and expert pathological reasoning. Ultimately, the proposed framework establishes a robust foundational step toward automated clinical support, though rigorous prospective, multi-center validations remain necessary to facilitate safe real-world deployment.