PDMR

KO | EN

gitlite — search

#rag #ruby #mcp #android #gpt #chatbot #database #nextjs #claude-code #reactjs

(★ 9)

[CVPR 2026] Prospective Dynamic 3D MRI Reconstruction via Latent-Space Motion Tracking from Single Measurement

File Explorer

ZIP Download

- dvf.pdf
- dvf.png
- inference_time.pdf
- inference_time.png
- pca_latent.pdf
- pca_latent.png
- pipeline-v3.pdf
- pipeline-v3.png
- teaser_demo.pdf
- teaser_demo.png
- video-coronal.gif
- video-coronal.mp4
- video-sagittal.gif
- video-sagittal.mp4
.gitignore
Prospective_Recon.pdf
README.md

# Use via CDN

jsDelivr

jsDelivr serves any public GitHub repository as a CDN with zero setup. Pick a version and a file to get a ready-to-paste link and snippet.

Version

File

Link

Example

// repository documentation

Was this content helpful?

★ 0 (0 ratings)

<div align="center">

## Prospective Dynamic 3D MRI Reconstruction via Latent-Space Motion Tracking from Single Measurement

**CVPR 2026** | [Paper PDF](Prospective_Recon.pdf)

This is the official project page of our work **"Prospective Dynamic 3D MRI Reconstruction via Latent-Space Motion Tracking from Single Measurement"**.

<a href="https://maopaom.github.io/">Lixuan Chen</a>1&nbsp;&nbsp;
 <a href="https://znanliu.github.io/">Zhongnan Liu</a>1&nbsp;&nbsp;
 <a href="https://experts.umich.edu/6748-jesse-hamilton">Jesse Hamilton</a>1&nbsp;&nbsp;
 <a href="https://medschool.umich.edu/profile/4905/james-m-balter">James M. Balter</a>1&nbsp;&nbsp;
 
 <a href="https://jjparkcv.github.io/">Jeong Joon Park</a>1✉️&nbsp;&nbsp;
 <a href="https://liyueshen.engin.umich.edu/">Liyue Shen</a>1✉️&nbsp;&nbsp;

1University of Michigan

</div>

## Overview

Prospective reconstruction is crucial for clinical applications such as MRI-guided radiotherapy, where the system must reconstruct the current anatomy and estimate motion from measurements acquired within the current latency window. This setting is challenging because each online update has only an ultra-sparse single measurement and an instant runtime requirement.

**PDMR** learns a patient-specific latent manifold of deformation vector fields offline, then performs online prospective reconstruction by optimizing only a low-dimensional latent vector. A geometry-aware tri-plane mapping network decodes the latent code into fine-grained 3D motion fields, enabling fast, motion-aware reconstruction from single-shot measurements.

![Teaser](https://raw.githubusercontent.com/maopaom/PDMR/main/Fig/teaser_demo.png)

## Method

PDMR has two stages:

- **Offline manifold learning:** learn a compact motion manifold and DVF mapping network from time-continuous sparse measurements.
- **Online prospective reconstruction:** freeze the mapping network and rapidly adapt only the latent vector for each new instantaneous measurement.

![Pipeline](https://raw.githubusercontent.com/maopaom/PDMR/main/Fig/pipeline-v3.png)

## Dynamic Visualizations

<table>
 <tr>
 <td align="center" width="50%">
 Coronal View
 </td>
 <td align="center" width="50%">
 Sagittal View
 </td>
 </tr>
 <tr>
 <td align="center" width="50%">
 <img src="Fig/video-coronal.gif" alt="Coronal dynamic visualization" width="100%">
 </td>
 <td align="center" width="50%">
 <img src="Fig/video-sagittal.gif" alt="Sagittal dynamic visualization" width="100%">
 </td>
 </tr>
</table>

> GIFs are generated from the original videos: [coronal MP4](Fig/video-coronal.mp4), [sagittal MP4](Fig/video-sagittal.mp4).

## Results

### Motion Field Visualization

The estimated DVFs capture respiratory dynamics while preserving anatomically static regions.

![DVF visualization](https://raw.githubusercontent.com/maopaom/PDMR/main/Fig/dvf.png)

### Latent Space Analysis

The dominant component of the learned latent vectors follows the respiratory motion signal across retrospective learning and prospective reconstruction settings.

![Latent PCA analysis](https://raw.githubusercontent.com/maopaom/PDMR/main/Fig/pca_latent.png)

### Runtime and Reconstruction Quality

PDMR achieves strong reconstruction fidelity with fast latent-space adaptation.

![Inference time](https://raw.githubusercontent.com/maopaom/PDMR/main/Fig/inference_time.png)

## Citation

```bibtex
@InProceedings{Chen_2026_CVPR,
    author    = {Chen, Lixuan and Liu, Zhongnan and Hamilton, Jesse and Balter, James M. and Park, Jeong Joon and Shen, Liyue},
    title     = {Prospective Dynamic 3D MRI Reconstruction via Latent-Space Motion Tracking from Single Measurement},
    booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
    month     = {June},
    year      = {2026},
    pages     = {5627-5636}
}
```

PDMR

File Explorer

# Use via CDN

Was this content helpful?

Recent Feedback

Download README