Scholarlink

<p>Accurately recognizing complex graphic patterns remains a critical challenge in advanced computer vision applications, especially in domains that require high-accuracy structural analysis, such as tumor-focused medical imaging. Whole-slide histopathology images of tumors often exhibit highly irregular, multiscale, and sparsely distributed morphological patterns. These are difficult to model using conventional transformers. While standard transformers have the following advantages over convolutional networks, such as improved global context modeling, their fixed attention patterns and high computational cost limit their ability to model the dynamic, content-dependent structures characteristic of pathological tumor tissue. This work presents HieraFocus, a spatially adaptive transformer framework specifically designed to address such limitations. It has two subblocks: Spatially Adaptive Transformer (SAT) and Hierarchical Multiscale Feature Refinement (HFR). First, robust morphological and structural features are extracted using FCN. SAT then selectively focuses on tumor regions of interest via deformable, content-aware sampling. This focused representation is then refined by the proposed hierarchical feature refinement mechanism and forwarded into the classification head for precise tumor discrimination while efficiently modeling the long-range dependencies. Experimental results from complex pattern benchmarks, including Camelyon16 whole-slide histopathology, show that HieraFocus outperforms standard Vision Transformers and hierarchical models, providing 5.2% improvement in tumor-focused medical image analysis accuracy while reducing computational complexity by 28%. These results strongly support HieraFocus as a tool for fast and accurate advanced complex pattern recognition in medical tumor analysis and other related visual fields.</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

HieraFocus: adaptive transformer-based recognition of complex graphic patterns in medical tumor imaging

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年3月24日

查看全文
  • Danylo Borovyk,
  • Khadija Alhumaid,
  • Denys Marichev,
  • Serhii Leoshchenko,
  • Andrii Oliinyk,
  • Jamil Abedalrahim Jamil Alsayaydeh

摘要

Accurately recognizing complex graphic patterns remains a critical challenge in advanced computer vision applications, especially in domains that require high-accuracy structural analysis, such as tumor-focused medical imaging. Whole-slide histopathology images of tumors often exhibit highly irregular, multiscale, and sparsely distributed morphological patterns. These are difficult to model using conventional transformers. While standard transformers have the following advantages over convolutional networks, such as improved global context modeling, their fixed attention patterns and high computational cost limit their ability to model the dynamic, content-dependent structures characteristic of pathological tumor tissue. This work presents HieraFocus, a spatially adaptive transformer framework specifically designed to address such limitations. It has two subblocks: Spatially Adaptive Transformer (SAT) and Hierarchical Multiscale Feature Refinement (HFR). First, robust morphological and structural features are extracted using FCN. SAT then selectively focuses on tumor regions of interest via deformable, content-aware sampling. This focused representation is then refined by the proposed hierarchical feature refinement mechanism and forwarded into the classification head for precise tumor discrimination while efficiently modeling the long-range dependencies. Experimental results from complex pattern benchmarks, including Camelyon16 whole-slide histopathology, show that HieraFocus outperforms standard Vision Transformers and hierarchical models, providing 5.2% improvement in tumor-focused medical image analysis accuracy while reducing computational complexity by 28%. These results strongly support HieraFocus as a tool for fast and accurate advanced complex pattern recognition in medical tumor analysis and other related visual fields.