Scholarlink

<p>In this paper, we implement a vision graph neural network (ViG) architecture to segment microstructures in X-ray computed tomography 3D data. Our ViG architecture is first trained on a synthetic augmented dataset, and then fine-tuned on experimental data to obtain an improved segmentation. Successively, we assess the accuracy of the segmentation on manually-labeled experimental slices. We exemplarily use the approach on a complex microstructure: a metal matrix composite, reinforced with two ceramic phases, intermetallic inclusions and a silicon network, in order to show the generality of our method. ViG model proves to be more efficient than U-Nets in adapting to new data when fine-tuned on a small portion of the experimental data. The fine-tuned ViG shows comparable performance to U-Nets, while largely reducing the number of trainable parameters, with the potential of greater adaptability and efficiency.</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

Vision GNN (ViG) architecture for a fine-tuned segmentation of a complex Al–Si metal matrix composite XCT volume

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2025年4月23日

查看全文
  • M. Lapenna,
  • A. Tsamos,
  • F. Faglioni,
  • R. Fioresi,
  • F. Zanchetta,
  • G. Bruno

摘要

In this paper, we implement a vision graph neural network (ViG) architecture to segment microstructures in X-ray computed tomography 3D data. Our ViG architecture is first trained on a synthetic augmented dataset, and then fine-tuned on experimental data to obtain an improved segmentation. Successively, we assess the accuracy of the segmentation on manually-labeled experimental slices. We exemplarily use the approach on a complex microstructure: a metal matrix composite, reinforced with two ceramic phases, intermetallic inclusions and a silicon network, in order to show the generality of our method. ViG model proves to be more efficient than U-Nets in adapting to new data when fine-tuned on a small portion of the experimental data. The fine-tuned ViG shows comparable performance to U-Nets, while largely reducing the number of trainable parameters, with the potential of greater adaptability and efficiency.