Scholarlink

We present 3D MTransINR, a novel model for volumetric MRI modality translation based on Implicit Neural Representations (INRs). In clinical workflows, acquiring a complete set of MRI modalities is often impractical, motivating methods that can infer multiple complementary missing sequences from available ones. Our model tackles this predictive synthesis task by translating one or more source modalities into one or more targets within a unified, resolution-independent framework. 3D MTransINR combines a shared multilayer perceptron (MLP) with voxel-wise bias modulation generated by a 3D U-Net, enabling anatomically consistent synthesis. We extend prior INR-based methods to full volumetric data and adversarial training, and evaluate our approach on the ProstateX and BraTS datasets. On ProstateX, 3D MTransINR outperforms a 3D Pix2Pix baseline in PSNR and SSIM across all targets and resolutions, showing strong robustness to resolution changes. On BraTS, it preserves finer structural details but is more sensitive to modality contrast, highlighting the challenges of generalisation across anatomies and intensity profiles.

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 MTransINR: a 3D Modality Translation Model Based on Implicit Neural Representations

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Maria Ysern,
  • Veronica Vilaplana

摘要

We present 3D MTransINR, a novel model for volumetric MRI modality translation based on Implicit Neural Representations (INRs). In clinical workflows, acquiring a complete set of MRI modalities is often impractical, motivating methods that can infer multiple complementary missing sequences from available ones. Our model tackles this predictive synthesis task by translating one or more source modalities into one or more targets within a unified, resolution-independent framework. 3D MTransINR combines a shared multilayer perceptron (MLP) with voxel-wise bias modulation generated by a 3D U-Net, enabling anatomically consistent synthesis. We extend prior INR-based methods to full volumetric data and adversarial training, and evaluate our approach on the ProstateX and BraTS datasets. On ProstateX, 3D MTransINR outperforms a 3D Pix2Pix baseline in PSNR and SSIM across all targets and resolutions, showing strong robustness to resolution changes. On BraTS, it preserves finer structural details but is more sensitive to modality contrast, highlighting the challenges of generalisation across anatomies and intensity profiles.