Scholarlink

<p>Non rigid image registration plays a vital role in medical image analysis and computer vision applications. It provides accurate alignment of deformable structures. However, conventional methods often face challenges in achieving both accuracy and robustness, particularly in detailed regions. Recent deep learning (DL) based approaches have demonstrated potential but frequently rely on intensity based loss functions, which may result in misalignment in regions with high feature density. This paper provides a DL approach, leveraging U-Net, for deformation field prediction. with a modified loss function, incorporating data similarity, deep feature, and smoothness constraints, taking advantage of deep features extracted from a pretrained visual geometry group (VGG16) model. Additionally, a spatial transformer network (STN) for image warping and a morphological refinement step using dilation and erosion are employed to further enhance registration accuracy. A rigorous evaluation of the proposed method was conducted on three diverse datasets, Retinal Fundus Images (FIRE), BraTS2019, and LEVIR-CD, demonstrating consistent and competitive registration performance in terms of mean square error (MSE) and structural similarity index measure (SSIM). Results achieved a higher structural similarity of 0.9779 and lower registration error of 0.0011 on FIRE dataset, and maintains strong generalization capabilities on BraTS2019, MSE of 0.0052 and SSIM of 0.9149, and LEVIR-CD, MSE of 0.0112 and SSIM of 0.9383, datasets. These results confirm the effectiveness of the proposed approach and highlight the benefits of incorporating deep feature constraints in non rigid registration, thereby paving the way for novel clinical and computer vision applications.</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

Deformation field prediction based on a modified loss function with U-Net for non rigid image registration

  • 期刊论文

  • 发表日期：

    2025年7月9日

查看全文
  • Omaima El Bahi,
  • Youssef Qaraai,
  • Ahmad El Allaoui

摘要

Non rigid image registration plays a vital role in medical image analysis and computer vision applications. It provides accurate alignment of deformable structures. However, conventional methods often face challenges in achieving both accuracy and robustness, particularly in detailed regions. Recent deep learning (DL) based approaches have demonstrated potential but frequently rely on intensity based loss functions, which may result in misalignment in regions with high feature density. This paper provides a DL approach, leveraging U-Net, for deformation field prediction. with a modified loss function, incorporating data similarity, deep feature, and smoothness constraints, taking advantage of deep features extracted from a pretrained visual geometry group (VGG16) model. Additionally, a spatial transformer network (STN) for image warping and a morphological refinement step using dilation and erosion are employed to further enhance registration accuracy. A rigorous evaluation of the proposed method was conducted on three diverse datasets, Retinal Fundus Images (FIRE), BraTS2019, and LEVIR-CD, demonstrating consistent and competitive registration performance in terms of mean square error (MSE) and structural similarity index measure (SSIM). Results achieved a higher structural similarity of 0.9779 and lower registration error of 0.0011 on FIRE dataset, and maintains strong generalization capabilities on BraTS2019, MSE of 0.0052 and SSIM of 0.9149, and LEVIR-CD, MSE of 0.0112 and SSIM of 0.9383, datasets. These results confirm the effectiveness of the proposed approach and highlight the benefits of incorporating deep feature constraints in non rigid registration, thereby paving the way for novel clinical and computer vision applications.