Scholarlink

<p>Skull structure defects negatively affect the involved patients in both health and mental conditions. The skull part completion was mostly conducted by pixel prediction on medical images or by skull shape mirroring from existing parts. The geometrical characteristics of the third dimension are lacking in two-dimensional image spaces, and the human skull structures are not symmetrical in three-dimensional spaces. Direct prediction of the missing skull part based on the existing part of the skull is a promising concept for enhancing cranial reconstruction accuracy. We investigated a novel procedure for predicting the skull’s missing part based on skull shape learning. The missing part’s shape was predicted from the existing part’s shape based on the shape relation between the missing and existing parts. The skull shape prediction was conducted on several shapes and locations of missing parts on the skull from our synthetic database. The skull shape prediction errors were dependent on the sizes and shapes of the missing parts. When the areas of missing parts were 1,439.91 ± 675.29&#xa0;mm<sup>2</sup>, 5,786.94 ± 1,242.28&#xa0;mm<sup>2</sup>, and 12,100.78 ± 5,219.74 mm<sup>2</sup>, the mean errors were 0.29&#xa0;mm, 0.42&#xa0;mm, and 0.54&#xa0;mm, respectively, with random shapes. When the areas of missing parts were 742.76 ± 250.55&#xa0;mm<sup>2</sup>, 2,683.04 ± 587.99&#xa0;mm<sup>2</sup>, and 11,335.18 ± 3,035.69&#xa0;mm<sup>2</sup>, the mean errors were 0.21&#xa0;mm, 0.36&#xa0;mm, and 0.51&#xa0;mm, respectively, for the squared shapes. These errors were all within the good accuracy in the cranial reconstruction application. The investigated method will be implemented into a computer-aided system for automatically predicting and printing out the skull’s missing structures. This study has three main contributions: (1) a novel procedure for classifying missing and existing parts from a defective skull, (2) a concept of predicting skull missing parts from the existing parts based on their shape relation, and (3) a large dataset of 757 skull structures with their pre-defined features and unified shapes support further researches. The source codes accompanied by a benchmark dataset can be accessed via the link: <a href="https://doi.org/10.5281/zenodo.12729483">https://doi.org/10.5281/zenodo.12729483</a>.</p> Graphic 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

Novel shape learning and prediction for skull part completion and reconstruction

  • 期刊论文

  • 发表日期：

    2025年4月23日

查看全文
  • Tan-Nhu Nguyen,
  • Phong-Phu Vo,
  • Vi-Do Tran,
  • Hoai-Nam Nguyen,
  • Hoang-Anh Pham,
  • Hong-An Le-Ngoc,
  • Thi-Tuong-Vi Nguyen,
  • Khanh-Linh Huynh,
  • Ngoc-Bich Le,
  • Thi-Hiep Nguyen,
  • Tien-Tuan Dao

摘要

Skull structure defects negatively affect the involved patients in both health and mental conditions. The skull part completion was mostly conducted by pixel prediction on medical images or by skull shape mirroring from existing parts. The geometrical characteristics of the third dimension are lacking in two-dimensional image spaces, and the human skull structures are not symmetrical in three-dimensional spaces. Direct prediction of the missing skull part based on the existing part of the skull is a promising concept for enhancing cranial reconstruction accuracy. We investigated a novel procedure for predicting the skull’s missing part based on skull shape learning. The missing part’s shape was predicted from the existing part’s shape based on the shape relation between the missing and existing parts. The skull shape prediction was conducted on several shapes and locations of missing parts on the skull from our synthetic database. The skull shape prediction errors were dependent on the sizes and shapes of the missing parts. When the areas of missing parts were 1,439.91 ± 675.29 mm2, 5,786.94 ± 1,242.28 mm2, and 12,100.78 ± 5,219.74 mm2, the mean errors were 0.29 mm, 0.42 mm, and 0.54 mm, respectively, with random shapes. When the areas of missing parts were 742.76 ± 250.55 mm2, 2,683.04 ± 587.99 mm2, and 11,335.18 ± 3,035.69 mm2, the mean errors were 0.21 mm, 0.36 mm, and 0.51 mm, respectively, for the squared shapes. These errors were all within the good accuracy in the cranial reconstruction application. The investigated method will be implemented into a computer-aided system for automatically predicting and printing out the skull’s missing structures. This study has three main contributions: (1) a novel procedure for classifying missing and existing parts from a defective skull, (2) a concept of predicting skull missing parts from the existing parts based on their shape relation, and (3) a large dataset of 757 skull structures with their pre-defined features and unified shapes support further researches. The source codes accompanied by a benchmark dataset can be accessed via the link: https://doi.org/10.5281/zenodo.12729483.

Graphic abstract