Scholarlink

<p>Accurate classification of cervical cytology images remains challenging because of subtle morphological differences, staining variability, and visual similarity among cell categories. Although deep learning methods have achieved strong classification performance, many existing approaches provide limited transparency regarding the image regions and morphological features that influence their predictions. This study presents an interpretable hybrid classification framework that combines GoogLeNet and MobileNet with a Vision Transformer (ViT) to extract multi-scale local morphology, fine-grained texture, and global contextual information from cervical cytology images. The framework incorporates a biomarker-guided Grad-CAM module, a neuro-symbolic inference component, and an ontology-aware semantic consistency module. The proposed Cervical Cytology Biomarker Ontology (CCBO) was constructed by adapting established principles of biomedical semantic representation and published cervical cytomorphological descriptions. The ontology represents hierarchical cell categories, biomarker associations, semantic compatibility constraints, and exclusion relationships to support structured verification of model outputs. Fuzzy IF–THEN rules were used to assess neural predictions according to predefined combinations of morphology-related biomarkers. In addition, an IoU-based Grad-CAM localization framework was introduced to quantify the spatial agreement between activation regions and morphology-guided biomarker reference masks. This multi-stage design extends the framework beyond conventional black-box classification by linking visual attribution, fuzzy rule evaluation, and ontology-based consistency assessment. Experimental results showed that the proposed hybrid model achieved an average AUC of 99.23%, sensitivity of 98.86%, and accuracy of 98.80% on the SIPaKMeD dataset. Illustrative case analyses further demonstrated the framework’s ability to produce structured explanations. For dyskeratotic samples, the Grad-CAM maps emphasized keratinization-related cytoplasmic regions and hyperchromatic nuclear areas, while the neuro-symbolic component identified agreement with the predefined dyskeratotic biomarker rules.</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

Interpretable hybrid CNN–ViT framework with neuro-symbolic clinical decision support for early cervical cancer diagnosis

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年8月4日

查看全文
  • Hanan T. Halawani,
  • Ebrahim Mohammed Senan,
  • Esam Mohammed Asem Othman,
  • M. Attique Khan,
  • Yousef Asiri,
  • Awad Alyousef,
  • Suliman Mohamed Fati

摘要

Accurate classification of cervical cytology images remains challenging because of subtle morphological differences, staining variability, and visual similarity among cell categories. Although deep learning methods have achieved strong classification performance, many existing approaches provide limited transparency regarding the image regions and morphological features that influence their predictions. This study presents an interpretable hybrid classification framework that combines GoogLeNet and MobileNet with a Vision Transformer (ViT) to extract multi-scale local morphology, fine-grained texture, and global contextual information from cervical cytology images. The framework incorporates a biomarker-guided Grad-CAM module, a neuro-symbolic inference component, and an ontology-aware semantic consistency module. The proposed Cervical Cytology Biomarker Ontology (CCBO) was constructed by adapting established principles of biomedical semantic representation and published cervical cytomorphological descriptions. The ontology represents hierarchical cell categories, biomarker associations, semantic compatibility constraints, and exclusion relationships to support structured verification of model outputs. Fuzzy IF–THEN rules were used to assess neural predictions according to predefined combinations of morphology-related biomarkers. In addition, an IoU-based Grad-CAM localization framework was introduced to quantify the spatial agreement between activation regions and morphology-guided biomarker reference masks. This multi-stage design extends the framework beyond conventional black-box classification by linking visual attribution, fuzzy rule evaluation, and ontology-based consistency assessment. Experimental results showed that the proposed hybrid model achieved an average AUC of 99.23%, sensitivity of 98.86%, and accuracy of 98.80% on the SIPaKMeD dataset. Illustrative case analyses further demonstrated the framework’s ability to produce structured explanations. For dyskeratotic samples, the Grad-CAM maps emphasized keratinization-related cytoplasmic regions and hyperchromatic nuclear areas, while the neuro-symbolic component identified agreement with the predefined dyskeratotic biomarker rules.