Accelerating cardiac diffusion tensor imaging with deep learning-based tensor de-noising and breath hold reduction. A step towards improved efficiency and clinical feasibility

Michael Tänzer, Andrew D. Scott, Zohya Khalique, Maria Molto, Ramyah Rajakulasingam, Ranil de Silva, Dudley J. Pennell, Pedro Ferreira, Guang Yang, Daniel Rueckert, Sònia Nielles‐VallespinJournal of Cardiovascular Magnetic Resonance 2025 · DOI

TL;DR

This paper introduces a novel deep learning framework, WGANUformer, that significantly accelerates cardiac diffusion tensor imaging (cDTI) by reducing scan times up to four-fold while maintaining diagnostic accuracy. By shifting de-noising from image to tensor space and employing advanced Vision Transformer architectures, the method enables faster, more efficient cardiac microstructure analysis.

What Problem Does This Paper Address?

Cardiac diffusion tensor imaging (cDTI) is a powerful, non-invasive technique for visualizing the intricate microstructure of the heart. However, acquiring high-quality cDTI data currently requires multiple breath-holds, leading to lengthy scan times (up to 192 minutes for a full heart scan). This prolonged acquisition makes it challenging for patients, especially those with cardiac conditions, and limits the widespread clinical adoption of cDTI. The core problem is reducing the number of required breath-holds to make cDTI more clinically feasible without sacrificing image quality or diagnostic accuracy.

Methods

The authors propose a novel deep learning framework called WGANUformer (WGUF), which introduces several key innovations:

  • Tensor-Space De-noising: Instead of de-noising individual diffusion-weighted images, WGUF operates directly in the tensor space. This paradigm shift is crucial for preserving the complex structural information encoded within the diffusion tensors.
  • Vision Transformer Ensemble: The framework utilizes an ensemble of Vision Transformer-based models, specifically optimized for processing tensor data. These models are trained using adversarial techniques (a Generative Adversarial Network, GAN) with a UFormer backbone, which is known for its efficiency in image restoration tasks.
  • Sophisticated Data Augmentation: A dynamic repetition selection strategy is employed to maximize the utility of limited training data. This involves sampling different subsets of repetitions from each scan, effectively tripling the training dataset and making the model more robust to variations in acquisition timing.
  • Ensemble Learning: To further improve robustness and reduce variability, the predictions from five independently trained WGUF models are averaged.

The framework builds upon a U-Net baseline and systematically evaluates the contribution of each component, including channel normalization and tensor-to-tensor training.

Datasets

The study utilized a substantial dataset comprising 744 cDTI datasets from 197 healthy volunteers and 547 patients with various cardiac conditions. This dataset was acquired using Siemens Skyra and Vida 3T MRI scanners. The data was split into training, validation, and testing sets using an 80:10:10 ratio, with patient-level splitting to prevent data leakage.

Key Insights and Ideas

  • Tensor-space de-noising is superior: Decoupling tensor computation from the de-noising process and operating directly on tensors significantly improves the preservation of structural information compared to image-based de-noising.
  • Vision Transformers excel at tensor data: The UFormer architecture, adapted for tensor processing, effectively handles the multi-channel nature of diffusion tensor data and the specific noise patterns encountered in cDTI.
  • Adversarial training enforces consistency: The GAN component helps enforce tensor consistency beyond simple error minimization, leading to perceptually higher quality maps.
  • Data augmentation is critical for limited datasets: Dynamic repetition sampling is a powerful technique to leverage existing data more effectively, especially in medical imaging where large, annotated datasets are scarce.
  • Ensembling boosts robustness: Averaging predictions from multiple models significantly enhances stability and reduces the impact of aleatoric uncertainty.
  • SIT cases reveal architectural robustness: The model's ability to accurately process data from patients with Situs Inversus Totalis (SIT), a rare anatomical variant not seen in training, highlights its strong generalization capabilities beyond learned anatomical priors.

Key Results

  • Scan Time Reduction: The proposed WGUFx5 model enables a four-fold reduction in scan time, achieving diagnostic quality with 3-5 breath-holds compared to the standard ~8 repetitions.
  • Improved Accuracy: The method achieved over 20% reduction in cDTI error metrics compared to the U-Net baseline, with fractional anisotropy (FA) errors improving from 0.09 to 0.07.
  • Preservation of Anatomical Features: The de-noised cDTI maps successfully preserved critical anatomical features, including infarct characterization and transmural cardiomyocyte orientation patterns.
  • Robustness in Pathology: While errors increased in clinical cases compared to healthy volunteers, WGUFx5 consistently outperformed the LLS baseline and showed better resilience to pathological variations, especially with reduced breath-holds.
  • Exceptional Performance in SIT: The model demonstrated remarkable accuracy in processing SIT cases, outperforming even classical LLS methods and highlighting its ability to generalize to unseen anatomical variations.

Limitations and Future Work

  • Under-represented pathologies: While the dataset is diverse, some cardiac conditions are under-represented, which could impact model performance for those specific pathologies.
  • Manual delineation: The current implementation requires manual delineation of the left ventricular myocardium, introducing an operator-dependent step.
  • LLS for tensor estimation: The study used a Linear Least Squares (LLS) method for initial tensor estimation. Future work could explore more robust techniques like Weighted Least Squares (WLS).
  • Ensemble vs. single network: The comparison involved an ensemble model (WGUFx5) against a single-network baseline (U-Net). While ensemble effects contribute to robustness, the architectural improvements are also significant.
  • Generalization to other sequences: The model was trained exclusively on STEAM acquisitions. Preliminary evaluation on a clinical spin-echo (SE) dataset showed a significant generalization challenge, indicating that further training with diverse acquisition sequences is needed to improve performance on unseen data types. Future work will focus on acquiring and integrating SE data to close this performance gap.

Summary updated

Abstract

This demonstrates substantial improvements in cDTI acquisition efficiency, achieving up to four-fold scan time reduction (3-5 breath-holds) while maintaining diagnostic accuracy across diverse cardiac pathologies.
BibTeX
@article{tanzer2025accelerating,
  author = {Michael Tänzer and Andrew D. Scott and Zohya Khalique and Maria Molto and Ramyah Rajakulasingam and Ranil de Silva and Dudley J. Pennell and Pedro Ferreira and Guang Yang and Daniel Rueckert and Sònia Nielles‐Vallespin},
  title = {{Accelerating cardiac diffusion tensor imaging with deep learning-based tensor de-noising and breath hold reduction. A step towards improved efficiency and clinical feasibility}},
  journal = {Journal of Cardiovascular Magnetic Resonance},
  year = {2025},
  doi = {10.1016/j.jocmr.2025.101971},
}
Figure