Scholarlink

<p>Automated analysis of cervical cytology images remains challenging due to substantial variability in cell morphology, staining quality, and imaging conditions, as well as the difficulty of simultaneously capturing global contextual relationships, fine-grained cellular morphology, and spatial dependencies among neighbouring cells. While recent deep learning approaches have explored convolutional neural networks, transformers, and hybrid architectures, many existing models rely on single attention mechanisms or static feature fusion, limiting their ability to jointly represent contextual and morphological cues. This study presents a hybrid deep learning framework that integrates Swin Transformer V2, EfficientNetV2-S, and a Dual-Attention BiLSTM module to enhance cervical cell classification performance. The Swin Transformer V2 component employs hierarchical shifted-window self-attention, enabling efficient modeling of long-range contextual relationships across cellular regions, while EfficientNetV2-S captures fine-grained nuclear and cytoplasmic morphological features through computationally efficient convolutional encoding. To complement these representations, a Dual-Attention BiLSTM module is introduced to model spatial–sequential dependencies among neighbouring feature embeddings while simultaneously emphasizing diagnostically relevant spatial regions and feature channels. This design enables more discriminative characterization of complex cytological structures compared with conventional hybrid architectures. The proposed hybrid model was systematically evaluated across five benchmark datasets SipakMed, Herlev, ISBI 2014, Mendeley LBC, and the recently introduced RIVA (2025) collectively comprising more than 22,800 annotated cervical cell instances from both conventional Pap smear and liquid-based cytology preparations. Extensive cross-dataset experiments demonstrate consistent improvements in classification accuracy and strong generalization across heterogeneous imaging conditions and acquisition protocols. In addition, interpretability analyses using attention visualization techniques indicate that the model focuses on clinically meaningful nuclear and cytoplasmic regions. Overall, the proposed framework provides a robust and interpretable approach for automated cervical cytology analysis and offers a scalable foundation for future AI-assisted cervical cancer screening and digital pathology applications.</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

Leveraging Enhanced Swin Transformer V2, EfficientNetV2, and Dual Attention BiLSTM for High Accuracy Cervical Cancer Diagnosis

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年7月8日

查看全文
  • Sandeep Kumar Thota,
  • Umesh Kumar Lilhore,
  • Sumitra Sureliya,
  • Anto Lourdu Xavier Raj Arockia Selvarathinam,
  • A Rasmi,
  • Arvind Panwar,
  • Lidia Gosy Tekeste,
  • Ehab Seif Ghith,
  • Heba G. Mohamed,
  • Sarita Simaiya

摘要

Automated analysis of cervical cytology images remains challenging due to substantial variability in cell morphology, staining quality, and imaging conditions, as well as the difficulty of simultaneously capturing global contextual relationships, fine-grained cellular morphology, and spatial dependencies among neighbouring cells. While recent deep learning approaches have explored convolutional neural networks, transformers, and hybrid architectures, many existing models rely on single attention mechanisms or static feature fusion, limiting their ability to jointly represent contextual and morphological cues. This study presents a hybrid deep learning framework that integrates Swin Transformer V2, EfficientNetV2-S, and a Dual-Attention BiLSTM module to enhance cervical cell classification performance. The Swin Transformer V2 component employs hierarchical shifted-window self-attention, enabling efficient modeling of long-range contextual relationships across cellular regions, while EfficientNetV2-S captures fine-grained nuclear and cytoplasmic morphological features through computationally efficient convolutional encoding. To complement these representations, a Dual-Attention BiLSTM module is introduced to model spatial–sequential dependencies among neighbouring feature embeddings while simultaneously emphasizing diagnostically relevant spatial regions and feature channels. This design enables more discriminative characterization of complex cytological structures compared with conventional hybrid architectures. The proposed hybrid model was systematically evaluated across five benchmark datasets SipakMed, Herlev, ISBI 2014, Mendeley LBC, and the recently introduced RIVA (2025) collectively comprising more than 22,800 annotated cervical cell instances from both conventional Pap smear and liquid-based cytology preparations. Extensive cross-dataset experiments demonstrate consistent improvements in classification accuracy and strong generalization across heterogeneous imaging conditions and acquisition protocols. In addition, interpretability analyses using attention visualization techniques indicate that the model focuses on clinically meaningful nuclear and cytoplasmic regions. Overall, the proposed framework provides a robust and interpretable approach for automated cervical cytology analysis and offers a scalable foundation for future AI-assisted cervical cancer screening and digital pathology applications.