Scholarlink

Inter-modality image registration is a challenging problem in medical image analysis, primarily due to absence of robust energy minimization objective functions that can handle variations across modalities. In this work, we introduce a novel framework that reformulates the inter-modality registration task as an intra-modality problem via image synthesis. Specifically, we jointly optimize a 3D diffusion-based deep generative model for image translation and a classical, greedy non-parametric diffeomorphic deformable image registration algorithm based on the log domain demons algorithm. Our approach does not require perfectly aligned training pairs: the registration component helps preserve the geometrical structure of the source modality during translation, while the synthesis model generates realistic target modality images that facilitate improved alignment. Unlike existing deep learning-based methods that rely on complex multi-term loss functions or adversarial training, our framework simplifies optimization by requiring only one loss function for each task (translation and registration) making it more stable and interpretable. We evaluate our method on a novel inter-modality dataset consisting of high-resolution postmortem (exvivo) T2w MRI scans (7T, 0.3 mm \(^{3}\) ) and corresponding antemortem (invivo) T1w MRI scans (1.5T/3T/7T, 0.7–1.0 mm \(^{3}\) ) from individuals with Alzheimer’s disease and related dementias (ADRD), where substantial atrophy is present between scan acquisitions. We demonstrate that our method achieves superior registration accuracy in the presence of large anatomical and contrast discrepancies. We open-source our codebase at the .

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

VIOLET: Volumetric Image registration via Optimization and Learning for Efficient image Translation

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Pulkit Khandelwal,
  • Michael Tran Duong,
  • Lisa M. Levorse,
  • Sydney A. Lim,
  • Nathaniel Gauthier,
  • Ved Shenoy,
  • Eunice Chung,
  • Amanda E. Denning,
  • Alejandra Bahena,
  • Winifred Trotman,
  • Christopher A. Olm,
  • Hamsanandini Radhakrishnan,
  • Ranjit Ittyerah,
  • Karthik Prabhakaran,
  • Gabor Mizsei,
  • Theresa Schuck,
  • John Robinson,
  • Daniel T. Ohm,
  • Jeffrey S. Phillips,
  • John A. Detre,
  • Edward B. Lee,
  • David J. Irwin,
  • Corey T. McMillan,
  • M. Dylan Tisdall,
  • Sandhitsu R. Das,
  • David A. Wolk,
  • Paul A. Yushkevich

摘要

Inter-modality image registration is a challenging problem in medical image analysis, primarily due to absence of robust energy minimization objective functions that can handle variations across modalities. In this work, we introduce a novel framework that reformulates the inter-modality registration task as an intra-modality problem via image synthesis. Specifically, we jointly optimize a 3D diffusion-based deep generative model for image translation and a classical, greedy non-parametric diffeomorphic deformable image registration algorithm based on the log domain demons algorithm. Our approach does not require perfectly aligned training pairs: the registration component helps preserve the geometrical structure of the source modality during translation, while the synthesis model generates realistic target modality images that facilitate improved alignment. Unlike existing deep learning-based methods that rely on complex multi-term loss functions or adversarial training, our framework simplifies optimization by requiring only one loss function for each task (translation and registration) making it more stable and interpretable. We evaluate our method on a novel inter-modality dataset consisting of high-resolution postmortem (exvivo) T2w MRI scans (7T, 0.3 mm \(^{3}\) ) and corresponding antemortem (invivo) T1w MRI scans (1.5T/3T/7T, 0.7–1.0 mm \(^{3}\) ) from individuals with Alzheimer’s disease and related dementias (ADRD), where substantial atrophy is present between scan acquisitions. We demonstrate that our method achieves superior registration accuracy in the presence of large anatomical and contrast discrepancies. We open-source our codebase at the .