Scholarlink

<p>In deep learning-driven 3D medical image segmentation, designing deep and complex network structures has become the trend. This enhances segmentation accuracy; however, it results in computational complexity, increased parameters, and unnecessary dependencies. Facing the challenges of complex structures and labeling difficulties in 3D medical images, we propose a novel lightweight semi-supervised learning framework, namely NLNet. It aims to promote the green and sustainable development of deep learning-driven medical image segmentation. For network lightweight in semi-supervised learning, we explore the essential mechanism of lightweight and seek the optimal balance between model performance and lightweight architecture. In the superficial layer, efficient and lightweight convolutional modules (Lite3DConvBlock) and a lightweight encoder (Lite3DEncoder) architecture are designed to accurately capture the basic edge, color, and texture information of 3D medical images. In the middle layer, we design a deep kernel attention convolutional module (DKAC) and mine complex features such as shapes and patterns. In the deep layer, a lightweight decoder (Lite3DDecoder) is designed to efficiently differentiate high-level structure from background information. We conduct an&#xa0;experimental assessment of the proposed model&#xa0;against mainstream&#xa0;state-of-the-art methods on left atrial (MRI) and pancreas (CT) datasets. The results indicate that our model leads in MACs, FLOPs, Params and File Size metrics. Meanwhile, the Dice, HD95, and ASD scores of our model achieve a better performance.</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

Advancing green development in 3D medical image segmentation: a novel lightweight, semi-supervised framework

  • 期刊论文

  • 发表日期：

    2025年9月12日

查看全文
  • Lei Chen,
  • Yikai Zhao,
  • Dongxu Yang,
  • Jieru Hou,
  • Wenhao Liu

摘要

In deep learning-driven 3D medical image segmentation, designing deep and complex network structures has become the trend. This enhances segmentation accuracy; however, it results in computational complexity, increased parameters, and unnecessary dependencies. Facing the challenges of complex structures and labeling difficulties in 3D medical images, we propose a novel lightweight semi-supervised learning framework, namely NLNet. It aims to promote the green and sustainable development of deep learning-driven medical image segmentation. For network lightweight in semi-supervised learning, we explore the essential mechanism of lightweight and seek the optimal balance between model performance and lightweight architecture. In the superficial layer, efficient and lightweight convolutional modules (Lite3DConvBlock) and a lightweight encoder (Lite3DEncoder) architecture are designed to accurately capture the basic edge, color, and texture information of 3D medical images. In the middle layer, we design a deep kernel attention convolutional module (DKAC) and mine complex features such as shapes and patterns. In the deep layer, a lightweight decoder (Lite3DDecoder) is designed to efficiently differentiate high-level structure from background information. We conduct an experimental assessment of the proposed model against mainstream state-of-the-art methods on left atrial (MRI) and pancreas (CT) datasets. The results indicate that our model leads in MACs, FLOPs, Params and File Size metrics. Meanwhile, the Dice, HD95, and ASD scores of our model achieve a better performance.