Scholarlink

<b>Purpose</b> <p>Deformable medical image registration is a critical task for aligning anatomical structures across patients, enabling accurate analysis in clinical and research contexts. This process faces challenges such as learning from relatively small datasets, managing large anatomical deformations, handling multi-modal imaging, and dealing with noisy or limited annotations.</p> <b>Methods</b> <p>We propose VNet-ST, a deep learning framework combining a 3D VNet encoder-decoder architecture with Spatial Transformer and WARP modules to estimate smooth, voxel-wise displacement fields for 3D image registration. WARP applies the predicted deformation field to resample and align the moving image. The model is trained end-to-end using a composite loss function comprising weighted L1 loss on image intensities, Dice-Focal loss on anatomical labels, and diffusion loss to regularize the deformation field.</p> <b>Results</b> <p>We evaluate VNet-ST on three inter-patient registration datasets: OASIS, a small brain MRI dataset challenging due to small anatomical structures, Abdomen CT-CT, a small abdominal CT dataset characterized by large deformations and limited training samples and KITS’23 is a challenging kidney CT dataset. The performance of VNet-ST is competitive with other recent methods, achieving among the highest Dice coefficients of 0.9608 on OASIS, 0.7000 on Abdomen CT-CT, 0.5789 on KITS’23 and Tversky scores of 0.9632 on OASIS, 0.7111 on Abdomen CT-CT, 0.5785 on KITS’23. Additionally, VNet-ST exhibits strong recall and specificity, indicating robust alignment and anatomically plausible deformation fields.</p> <b>Conclusion</b> <p>These results demonstrate the potential of VNet-ST as a reliable tool for medical image registration, effectively addressing challenges posed by small datasets, large deformations, and small anatomical structures without compromising computational efficiency or registration accuracy.</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

3D Hybrid VNet and spatial transformer model for deformable medical image registration

  • 期刊论文

  • 发表日期：

    2026年2月9日

查看全文
  • Praveen M A,
  • Veera Bhadra Rao Donthamsetti,
  • Aravindh Mareedu,
  • Nithin Sai Manubolu,
  • Srilatha Chebrolu

摘要

Purpose

Deformable medical image registration is a critical task for aligning anatomical structures across patients, enabling accurate analysis in clinical and research contexts. This process faces challenges such as learning from relatively small datasets, managing large anatomical deformations, handling multi-modal imaging, and dealing with noisy or limited annotations.

Methods

We propose VNet-ST, a deep learning framework combining a 3D VNet encoder-decoder architecture with Spatial Transformer and WARP modules to estimate smooth, voxel-wise displacement fields for 3D image registration. WARP applies the predicted deformation field to resample and align the moving image. The model is trained end-to-end using a composite loss function comprising weighted L1 loss on image intensities, Dice-Focal loss on anatomical labels, and diffusion loss to regularize the deformation field.

Results

We evaluate VNet-ST on three inter-patient registration datasets: OASIS, a small brain MRI dataset challenging due to small anatomical structures, Abdomen CT-CT, a small abdominal CT dataset characterized by large deformations and limited training samples and KITS’23 is a challenging kidney CT dataset. The performance of VNet-ST is competitive with other recent methods, achieving among the highest Dice coefficients of 0.9608 on OASIS, 0.7000 on Abdomen CT-CT, 0.5789 on KITS’23 and Tversky scores of 0.9632 on OASIS, 0.7111 on Abdomen CT-CT, 0.5785 on KITS’23. Additionally, VNet-ST exhibits strong recall and specificity, indicating robust alignment and anatomically plausible deformation fields.

Conclusion

These results demonstrate the potential of VNet-ST as a reliable tool for medical image registration, effectively addressing challenges posed by small datasets, large deformations, and small anatomical structures without compromising computational efficiency or registration accuracy.