Scholarlink

<p>Skull part segmentation is of great interest in medical and forensic applications for analyzing, diagnosing, and reconstructing skull defects. Segmentation using medical imaging is currently limited to two-dimensional (2-D) images. Moreover, the segmentation process is time-consuming and less accurate. Thus, our objective was to segment skull parts directly from three-dimensional (3-D) shapes with a competitive benchmarking of current state-of-the-art deep learning approaches.&#xa0;We reconstructed 329 skull meshes from 329 Computed Tomography (CT) scans and then generated associated skull shapes. A skull shape template with pre-defined parts (i.e., segments of the subject’s skull shape) was then optimally deformed to the reconstructed skull meshes. We employed three advanced deep learning approaches, including geometric deep learning (GDL), generative adversarial network (GAN), and point transformer (PT), for optimally classifying the vertices on the skull parts. A 5-fold cross-validation procedure was conducted to select the optimal clustering concept with its appropriate parameters. The accuracy (ACC), intersection over union (IoU), and dice score (DSC) metrics were used for evaluating the training, testing, and validating results.&#xa0;After the cross-validation, the optimal approach was the GAN coupled with a Geometric Deep Network (GDN). The optimal IoU, DSC, and ACC were 95.79% ± 0.55%, 97.85% ± 0.29%, and 97.85% ± 0.29%, respectively.&#xa0;This study proposed and evaluated a novel procedure for segmenting skull parts directly from 3-D skull shapes. The competitive benchmarking allowed us to select the best approach for this complex issue. Obtained outcomes open a new concept for automatically detecting defective skull regions from a 3-D skull shape for skull-defect reconstruction.</p> Graphical abstract <p></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

A competitive benchmarking among geometric deep learning, generative adversarial network, and point transformer for 3-D skull part segmentation

  • 期刊论文

  • 发表日期：

    2025年11月2日

查看全文
  • Thi-Thuy-Duong Pham,
  • Duc-Phong Nguyen,
  • Hien Duyen Nguyen-Le,
  • Tan-Nhu Nguyen,
  • Tien-Tuan Dao

摘要

Skull part segmentation is of great interest in medical and forensic applications for analyzing, diagnosing, and reconstructing skull defects. Segmentation using medical imaging is currently limited to two-dimensional (2-D) images. Moreover, the segmentation process is time-consuming and less accurate. Thus, our objective was to segment skull parts directly from three-dimensional (3-D) shapes with a competitive benchmarking of current state-of-the-art deep learning approaches. We reconstructed 329 skull meshes from 329 Computed Tomography (CT) scans and then generated associated skull shapes. A skull shape template with pre-defined parts (i.e., segments of the subject’s skull shape) was then optimally deformed to the reconstructed skull meshes. We employed three advanced deep learning approaches, including geometric deep learning (GDL), generative adversarial network (GAN), and point transformer (PT), for optimally classifying the vertices on the skull parts. A 5-fold cross-validation procedure was conducted to select the optimal clustering concept with its appropriate parameters. The accuracy (ACC), intersection over union (IoU), and dice score (DSC) metrics were used for evaluating the training, testing, and validating results. After the cross-validation, the optimal approach was the GAN coupled with a Geometric Deep Network (GDN). The optimal IoU, DSC, and ACC were 95.79% ± 0.55%, 97.85% ± 0.29%, and 97.85% ± 0.29%, respectively. This study proposed and evaluated a novel procedure for segmenting skull parts directly from 3-D skull shapes. The competitive benchmarking allowed us to select the best approach for this complex issue. Obtained outcomes open a new concept for automatically detecting defective skull regions from a 3-D skull shape for skull-defect reconstruction.

Graphical abstract