Scholarlink

<p><i>Objective:</i> Cell classification is crucial for medical early diagnosis and personalized treatment. However, it remains a significant challenge to achieve a balance between classification accuracy and model lightweighting in microscope-based cell image classification. This study aims to optimize inference speed and reduce deployment complexity for this task. <i>Approach:</i> We propose a novel method based on the Vicinity Vision Transformer (VicinityViT) and channel-position attention (CPA), named Lightweight Channel Position Transformer (LCPT). Dynamic convolutions preserve classification accuracy, while fused channel and positional information enhances small-object detection. Contrastive learning-based normalization helps prevent attention mechanism collapse, and skip connections address gradient flow issues. Furthermore, the integration of Kolmogorov–Arnold Networks (KAN) reduces both computational complexity and parameter count. <i>Main results:</i> Experiments conducted on four public datasets yielded accuracies of 95.94% (BioMediTech), 96.17% (ICPR-HEp-2), 87.89% (Hematological Malignancy Bone Marrow Cytology Expert Annotation for a six-category classification), and 97.84% (White Blood Cell). It shows remarkable efficiency, with 0.03 GFLOPs and 0.91M parameters. <i>Significance:</i> Our approach significantly improves model lightweighting with minimal accuracy degradation, balancing efficiency and performance, outperforming state-of-the-art methods in compactness and versatility, demonstrating potential in efficient model deployment and sustainable AI development.</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

LCPT: A lightweight cell classification method for microscope images based on vicinityViT and channel-position attention

  • 期刊论文

  • 发表日期：

    2025年10月3日

查看全文
  • Panpan Wu,
  • Zhangda Liu,
  • Ziping Zhao,
  • Rui Guo,
  • Hengyong Yu

摘要

Objective: Cell classification is crucial for medical early diagnosis and personalized treatment. However, it remains a significant challenge to achieve a balance between classification accuracy and model lightweighting in microscope-based cell image classification. This study aims to optimize inference speed and reduce deployment complexity for this task. Approach: We propose a novel method based on the Vicinity Vision Transformer (VicinityViT) and channel-position attention (CPA), named Lightweight Channel Position Transformer (LCPT). Dynamic convolutions preserve classification accuracy, while fused channel and positional information enhances small-object detection. Contrastive learning-based normalization helps prevent attention mechanism collapse, and skip connections address gradient flow issues. Furthermore, the integration of Kolmogorov–Arnold Networks (KAN) reduces both computational complexity and parameter count. Main results: Experiments conducted on four public datasets yielded accuracies of 95.94% (BioMediTech), 96.17% (ICPR-HEp-2), 87.89% (Hematological Malignancy Bone Marrow Cytology Expert Annotation for a six-category classification), and 97.84% (White Blood Cell). It shows remarkable efficiency, with 0.03 GFLOPs and 0.91M parameters. Significance: Our approach significantly improves model lightweighting with minimal accuracy degradation, balancing efficiency and performance, outperforming state-of-the-art methods in compactness and versatility, demonstrating potential in efficient model deployment and sustainable AI development.