Related Experiment Video
Updated: Apr 14, 2026
11:36
Voxel Printing Anatomy: Design and Fabrication of Realistic, Presurgical Planning Models through Bitmap Printing
Published on: February 9, 2022
3.4K
Cyclic 2.5D Perceptual Loss for Cross-Modal 3D Medical Image Synthesis: T1w MRI to Tau PET
Junho Moon1, Symac Kim2, Haejun Chung1,2,3
1Department of Artificial Intelligence Semiconductor Engineering, Hanyang University, Seoul, South Korea.
Human Brain Mapping
|April 12, 2026
Summary
This study introduces a novel method for synthesizing 3D tau PET scans from MRI, improving Alzheimer's disease diagnosis. The technique enhances volumetric consistency and reduces scanner variability, aiding in early detection and treatment strategies for neurodegenerative diseases.
Area of Science:
- Neuroimaging
- Medical Image Synthesis
- Alzheimer's Disease and Related Dementias (ADRD)
Background:
- Positron emission tomography (PET) is crucial for diagnosing Alzheimer's disease and related dementias (ADRD), but its use is limited by cost and invasiveness.
- Modern diagnostic frameworks like the amyloid/tau/neurodegeneration (A/T/N) require multimodal biomarkers, posing challenges due to PET's accessibility.
- Medical image synthesis offers a way to generate missing imaging modalities, potentially overcoming these limitations.
Purpose of the Study:
- To develop a method for synthesizing 3D tau PET scans (specifically pseudo-[18F]flortaucipir SUVR maps) from structural T1-weighted MRI.
- To improve the accuracy and clinical utility of PET imaging by addressing limitations in generative model synthesis.
- To enhance volumetric consistency and reduce inter-scanner variability in synthesized PET data.
Main Methods:
- Proposed a cyclic 2.5D perceptual loss function to optimize synthesis across axial, coronal, and sagittal planes, improving volumetric consistency.
- Standardized PET SUVRs by scanner manufacturer to minimize inter-manufacturer variability and preserve high-uptake regions.
- Evaluated the approach on diverse generative architectures (U-Net, UNETR, SwinUNETR, CycleGAN, Pix2Pix) using ADRD cohort data.
Main Results:
- The cyclic 2.5D perceptual loss enhanced volumetric consistency in synthesized 3D tau PET scans.
- Standardization by scanner manufacturer improved the preservation of high-uptake regions and reduced variability.
- The method demonstrated high quantitative and qualitative performance across various generative models, achieving better agreement with actual PET scans in key Alzheimer's-related brain regions.
Conclusions:
- The proposed image synthesis approach effectively generates 3D pseudo-tau PET SUVR maps from MRI, addressing key limitations of current methods.
- This technique shows broad applicability across different generative frameworks and holds promise for improving multimodal biomarker assessment in ADRD.
- The publicly available code facilitates further research and clinical application of this advanced neuroimaging synthesis method.
