Scholarlink

<p>The global burden of cervical cancer remains substantial because patient prognosis depends significantly on correct cytological classification, together with early detection of the disease. Convolutional Neural Networks (CNNs) excel at extracting spatially localised morphological patterns yet struggle to model the long-range dependencies found in complex cellular structures. Vision Transformers (ViTs) demonstrate potential for encoding global context through self-attention yet struggle with detailed local feature extraction on low-resolution cytological datasets when used alone. We present ViDeCerviNet as a hybrid deep learning framework that merges DenseNet201 with ViT_base_patch16_128 through a parallel dual-branch system and uses a concatenation-based feature fusion approach to overcome existing architectural limitations. The model combines hierarchical convolutional feature maps with global self-attention representations to enable complete multi-scale feature extraction from Pap smear cytology images. Through evaluation on the SIPaKMeD benchmark dataset, which tested 11 pre-trained CNNs and 12 pre-trained Transformer variants along with 20 hybrid models, ViDeCerviNet achieved a top classification accuracy of 99.26% that surpassed all existing unimodal systems and previously tested hybrid architectures. The model shows strong generalisation capabilities across various cytological subtypes while maintaining high accuracy in diagnostic classification and precise categorisation of precancerous lesions like dyskeratotic, koilocytotic, and metaplastic cells. The hybridisation strategy demonstrates how composite models that combine spatial precision with global contextualisation enhance diagnostic capabilities for AI-assisted cervical cancer screening. The system’s minimal weight structure, modular composition and real-time inference pipeline compatibility make it ideal for implementation in clinical screening operations and point-of-care diagnostic settings.</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

ViDeCerviNet: a hybrid vision transformer–DenseNet framework for superior accuracy in cervical cancer diagnosis and categorisation

  • 期刊论文

  • 发表日期：

    2026年6月23日

查看全文
  • Amol Satsangi,
  • K. Srinivas,
  • A. Charan Kumari

摘要

The global burden of cervical cancer remains substantial because patient prognosis depends significantly on correct cytological classification, together with early detection of the disease. Convolutional Neural Networks (CNNs) excel at extracting spatially localised morphological patterns yet struggle to model the long-range dependencies found in complex cellular structures. Vision Transformers (ViTs) demonstrate potential for encoding global context through self-attention yet struggle with detailed local feature extraction on low-resolution cytological datasets when used alone. We present ViDeCerviNet as a hybrid deep learning framework that merges DenseNet201 with ViT_base_patch16_128 through a parallel dual-branch system and uses a concatenation-based feature fusion approach to overcome existing architectural limitations. The model combines hierarchical convolutional feature maps with global self-attention representations to enable complete multi-scale feature extraction from Pap smear cytology images. Through evaluation on the SIPaKMeD benchmark dataset, which tested 11 pre-trained CNNs and 12 pre-trained Transformer variants along with 20 hybrid models, ViDeCerviNet achieved a top classification accuracy of 99.26% that surpassed all existing unimodal systems and previously tested hybrid architectures. The model shows strong generalisation capabilities across various cytological subtypes while maintaining high accuracy in diagnostic classification and precise categorisation of precancerous lesions like dyskeratotic, koilocytotic, and metaplastic cells. The hybridisation strategy demonstrates how composite models that combine spatial precision with global contextualisation enhance diagnostic capabilities for AI-assisted cervical cancer screening. The system’s minimal weight structure, modular composition and real-time inference pipeline compatibility make it ideal for implementation in clinical screening operations and point-of-care diagnostic settings.