Structure-aware unsupervised tagged-to-cine MRI synthesis with self disentanglement

Xiaofeng Liu, Fangxu Xing, Jerry L. Prince, Maureen Stone, Georges El Fakhri, Jonghye Woo

Introduction

In cardiac and speech imaging, Tagged MRI is a vital technique that uses magnetic tags to track tissue deformation. However, these tags often obscure anatomical details. Cine MRI provides a clear view of anatomy but requires additional scan time and cost. Synthesizing Cine MRI from Tagged MRI data could streamline clinical workflows, but existing deep learning methods face two major hurdles: the scarcity of perfectly paired training data and the tendency of unsupervised models (like CycleGAN) to introduce "geometric distortions." These distortions can shift anatomical boundaries, making the images unreliable for medical diagnosis.

Proposed Methodology

To address these challenges, we propose a Structure-aware Unsupervised Tagged-to-Cine MRI Synthesis framework. The core innovation is a Structure Extraction Module (SEM) integrated into a standard CycleGAN architecture. This module enforces anatomical consistency without requiring manual labels or paired data.

The proposed framework featuring the CycleGAN architecture and the novel Structure Extraction Module for modality-invariant feature learning.

Self-Disentanglement via Patches

The SEM operates on the principle that while the "style" (tags vs. clear anatomy) changes between modalities, the underlying structure remains invariant. The module uses a feature extractor ($f$) and a classifier ($C$) in an adversarial game. By comparing local $3 \times 3$ patches, the extractor is trained to "confuse" the classifier, forcing it to ignore modality-specific features (the tags) and focus purely on anatomical structure. This ensures that the synthesized Cine MRI maintains the exact geometry of the input Tagged MRI.

Experimental Evaluation

The framework was evaluated using a dataset of 20 healthy subjects performing speech tasks. The training set consisted of 1,768 unpaired slices, while validation and testing were performed on 416 and 1,560 paired slices, respectively. Performance was measured using Structural Similarity (SSIM), Peak Signal-to-Noise Ratio (PSNR), and Inception Score (IS).

Results and Discussion

The proposed method significantly outperformed vanilla GAN and CycleGAN architectures across all quantitative metrics. Specifically, the model achieved an SSIM of 0.9237, compared to 0.9014 for CycleGAN, and a PSNR of 29.52, compared to 27.33.

Qualitative comparison of synthesis results. The proposed method (right) maintains sharp anatomical boundaries and tongue shape compared to the distortions seen in standard GAN and CycleGAN.

Qualitatively, the images produced by our method show much higher fidelity to the ground truth. While standard CycleGAN often produces blurry results or shifts anatomical boundaries (as seen in the red-boxed tongue regions), our approach preserves the sharp, consistent shape of the tongue as seen in the original Tagged input. This demonstrates that local patch-based disentanglement effectively penalizes the small-scale distortions that typically plague unsupervised image-to-image translation.

Conclusion

This work introduces a self-supervised "add-on" module that enables high-fidelity MRI synthesis without the need for paired data or manual segmentation. By focusing on modality-invariant structure, the model overcomes the geometric distortion issues common in previous unsupervised methods. Future research will explore the generalizability of this framework to other tasks, such as MRI-to-CT synthesis for radiation therapy planning.

Summary updated

BibTeX
@inproceedings{liu2022structure,
  author = {Xiaofeng Liu and Fangxu Xing and Jerry L. Prince and Maureen Stone and Georges El Fakhri and Jonghye Woo},
  title = {{Structure-aware unsupervised tagged-to-cine MRI synthesis with self disentanglement}},
  booktitle = {Medical Imaging: Image Processing},
  year = {2022},
  url = {https://openreview.net/forum?id=DDwlOoYqs4},
}
Figure