Scholarlink

<p>Precise segmentation is crucial for accurate computer-aided diagnosis systems, as segmenting the vertebral column from the medical images reduces noise and complexity, enabling more effective model training. However, identifying spinal discs and bones in medical images is challenging, especially when dealing with abnormalities. Although, Deep Convolutional Neural Networks (DCNNs) have shown potential in medical image segmentation, however, they often struggle with complex anatomical structures and data limitations, requiring significant computational resources, energy, and time. Current deep learning models for segmentation are typically highly parameterized, further increasing complexity. In response, this paper proposes a Lightweight deep learning framework, i.e., <i>LiTransUNet</i>, built on the top of TransUNet to segment Vertebral Columns from dog radiography samples. It uses depthwise separable and pointwise convolution along with a transformer to encode dataset samples instead of traditional CNNs. The proposed model is trained on radiographic images collected by the Department of Surgery and Radiology at the College of Veterinary Sciences, GBPUA&amp;T, Pantnagar. Experimental results show the proposed method yields comparable performance while significantly reducing trainable parameters, memory usage, energy consumption, and processing time.</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

LiTransUNet: a lightweight TransUNet for vertebral column segmentation

  • 期刊论文

  • 发表日期：

    2025年5月2日

查看全文
  • Krishna K. Agrawal,
  • Rajneesh Kumar,
  • Gautam Kumar

摘要

Precise segmentation is crucial for accurate computer-aided diagnosis systems, as segmenting the vertebral column from the medical images reduces noise and complexity, enabling more effective model training. However, identifying spinal discs and bones in medical images is challenging, especially when dealing with abnormalities. Although, Deep Convolutional Neural Networks (DCNNs) have shown potential in medical image segmentation, however, they often struggle with complex anatomical structures and data limitations, requiring significant computational resources, energy, and time. Current deep learning models for segmentation are typically highly parameterized, further increasing complexity. In response, this paper proposes a Lightweight deep learning framework, i.e., LiTransUNet, built on the top of TransUNet to segment Vertebral Columns from dog radiography samples. It uses depthwise separable and pointwise convolution along with a transformer to encode dataset samples instead of traditional CNNs. The proposed model is trained on radiographic images collected by the Department of Surgery and Radiology at the College of Veterinary Sciences, GBPUA&T, Pantnagar. Experimental results show the proposed method yields comparable performance while significantly reducing trainable parameters, memory usage, energy consumption, and processing time.