Scholarlink

<p>3D computed tomography (CT) image segmentation has been widely studied due to its important role in disease diagnosis and treatment. Since most existing methods rely on expensive manual annotations, self-supervised learning has been introduced to this task. MAE and its variants in 3D medical image analysis have achieved significant performance improvements, but challenges still exist. Firstly, indiscriminately reconstructing voxels leads to learning unimportant areas and redundant information in CT images. Secondly, failing to fully utilize the positional prior information related to fixed human structures in 3D CT images. To address these challenges, this paper proposes the GFPP-MAE model, which consists of the gradient-guided frequency re-construction module (GFRM), the absolute position prediction module (APPM) and the relative position prediction module (RPPM). GFRM reconstructs CT images in the frequency domain and utilizes gradient-guided weighted loss to focus on important edge areas, which helps to avoid learning redundant features and concentrate modeling capability. APPM and RPPM are used to enhance spatial structure perception. APPM learns global structure information by predicting the absolute position. RPPM understands local–global structure consistency by predicting the volume proportion of randomly cropped sub-volume in each base block. The experiments of abdominal multi-organ segmentation on the BTCV dataset and lung tumor segmentation on the MSD Lung dataset both demonstrate that the GFPP-MAE outperforms other state-of-the-art models. The experiment on the unseen AMOS dataset, which is excluded during pre-training, demonstrates the model’s superior generalization capability. The code is available in <a href="https://github.com/Dmitvna/GFPP-MAE.git">https://github.com/Dmitvna/GFPP-MAE.git</a>.</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

GFPP-MAE: gradient-guided frequency reconstruction and position predictions advance MAE for 3D CT image segmentation

  • 期刊论文

  • 发表日期：

    2026年2月3日

查看全文
  • Yuping Peng,
  • Xing Wu,
  • Xing Xiao,
  • Chengliang Wang,
  • Hongqian Wang

摘要

3D computed tomography (CT) image segmentation has been widely studied due to its important role in disease diagnosis and treatment. Since most existing methods rely on expensive manual annotations, self-supervised learning has been introduced to this task. MAE and its variants in 3D medical image analysis have achieved significant performance improvements, but challenges still exist. Firstly, indiscriminately reconstructing voxels leads to learning unimportant areas and redundant information in CT images. Secondly, failing to fully utilize the positional prior information related to fixed human structures in 3D CT images. To address these challenges, this paper proposes the GFPP-MAE model, which consists of the gradient-guided frequency re-construction module (GFRM), the absolute position prediction module (APPM) and the relative position prediction module (RPPM). GFRM reconstructs CT images in the frequency domain and utilizes gradient-guided weighted loss to focus on important edge areas, which helps to avoid learning redundant features and concentrate modeling capability. APPM and RPPM are used to enhance spatial structure perception. APPM learns global structure information by predicting the absolute position. RPPM understands local–global structure consistency by predicting the volume proportion of randomly cropped sub-volume in each base block. The experiments of abdominal multi-organ segmentation on the BTCV dataset and lung tumor segmentation on the MSD Lung dataset both demonstrate that the GFPP-MAE outperforms other state-of-the-art models. The experiment on the unseen AMOS dataset, which is excluded during pre-training, demonstrates the model’s superior generalization capability. The code is available in https://github.com/Dmitvna/GFPP-MAE.git.