Scholarlink

<p>Cervical cancer remains a leading cause of cancer-related mortality among women globally, with early detection through automated cytology screening critical for improving patient outcomes. This study presents HybridKANViTAE, a novel federated learning framework for privacy-preserving cervical cancer classification that synergistically integrates three complementary architectural branches: an Autoencoder-CNN pathway for morphology-aware feature extraction, a Vision Transformer for global contextual modeling, and a Kolmogorov-Arnold Network with learnable activation functions. A learnable weighted fusion mechanism dynamically combines outputs from all branches, adapting to heterogeneous data distributions encountered in federated settings. Comprehensive evaluation on two benchmark datasets (SIPaKMeD and APCData) demonstrates robust performance with 99.68% accuracy (MCC: 0.9960) and 91.69% accuracy (MCC: 0.8616) respectively, while federated training achieves performance within 0.12–0.55% of centralized approaches. Extensive ablation studies validate the contribution of each architectural component, with learnable fusion outperforming fixed fusion by 0.67–0.83 percentage points. Comprehensive explainable AI techniques including Grad-CAM, attention visualization, and branch contribution analysis ensure clinical interpretability by highlighting diagnostically relevant morphological features aligned with the Bethesda System criteria. The proposed framework addresses critical gaps in privacy-preserving collaborative learning for medical imaging, enabling multi-institutional model development while maintaining strict data confidentiality and regulatory compliance with HIPAA and GDPR. Results demonstrate that federated learning with sophisticated hybrid architectures provides a viable pathway for deploying AI-based cervical cancer screening systems in distributed clinical environments.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Federated learning-based cervical cancer classification using a novel hybrid KAN-ViT-autoencoder architecture

  • 期刊论文

  • 发表日期：

    2026年7月18日

查看全文
  • Mohammed Tawik,
  • Mohamed Heshmat,
  • Islam S. Fathi,
  • Ahmed S. Shaban,
  • Ghazi Shakah,
  • Warda M. Shaban

摘要

Cervical cancer remains a leading cause of cancer-related mortality among women globally, with early detection through automated cytology screening critical for improving patient outcomes. This study presents HybridKANViTAE, a novel federated learning framework for privacy-preserving cervical cancer classification that synergistically integrates three complementary architectural branches: an Autoencoder-CNN pathway for morphology-aware feature extraction, a Vision Transformer for global contextual modeling, and a Kolmogorov-Arnold Network with learnable activation functions. A learnable weighted fusion mechanism dynamically combines outputs from all branches, adapting to heterogeneous data distributions encountered in federated settings. Comprehensive evaluation on two benchmark datasets (SIPaKMeD and APCData) demonstrates robust performance with 99.68% accuracy (MCC: 0.9960) and 91.69% accuracy (MCC: 0.8616) respectively, while federated training achieves performance within 0.12–0.55% of centralized approaches. Extensive ablation studies validate the contribution of each architectural component, with learnable fusion outperforming fixed fusion by 0.67–0.83 percentage points. Comprehensive explainable AI techniques including Grad-CAM, attention visualization, and branch contribution analysis ensure clinical interpretability by highlighting diagnostically relevant morphological features aligned with the Bethesda System criteria. The proposed framework addresses critical gaps in privacy-preserving collaborative learning for medical imaging, enabling multi-institutional model development while maintaining strict data confidentiality and regulatory compliance with HIPAA and GDPR. Results demonstrate that federated learning with sophisticated hybrid architectures provides a viable pathway for deploying AI-based cervical cancer screening systems in distributed clinical environments.