Abstract
High-temporal-resolution cardiac MRI is essential for diagnosing heart disease but often suffers from long acquisition times or low frame rates. This paper presents a multi-task deep learning framework that synthesizes intermediate cardiac frames and performs automated left ventricle segmentation simultaneously. Using a multi-head multi-tail attention-based UNet, the model achieves high fidelity in both magnitude and phase image synthesis, significantly outperforming traditional interpolation methods.
Introduction
To diagnose cardiac diseases effectively, clinicians must observe myocardial motion in real-time. Three-directional myocardial velocity mapping (3Dir MVM) is the gold standard for this, providing anatomical magnitude images and phase images that encode tissue velocity. However, 3Dir MVM is constrained by a trade-off: long scans introduce breathing artifacts, while shorter scans result in low temporal resolution. This work addresses the challenge of "filling in the gaps" by synthesizing intermediate frames from low-frame-rate acquisitions using deep learning.
Methodology
The core of the proposed system is a multi-head multi-tail attention-based UNet architecture designed for multi-task learning.
Architecture
As shown in Figure 1, the model employs three separate encoders to process magnitude images, phase images, and myocardium segmentation masks. These features are concatenated and then processed by task-specific decoders. Attention gates help the model focus on relevant features, while skip connections preserve fine-grained spatial details during synthesis.
Conditional Synthesis and Loss Functions
To guide the network in generating specific frames, a "conditional feature map" is introduced. This map provides normalized time points ($k$) within the cardiac cycle, allowing the model to account for non-linear heart motion.
The training process utilizes custom loss functions to prioritize clinical relevance over background noise. A weighted Mean Absolute Error (MAE) forces the model to focus on the myocardium, while a combination of Dice loss and Boundary loss ensures sharp, accurate segmentation edges.
Results
The model was evaluated on an in-house dataset from the Royal Brompton Hospital, consisting of 30 healthy volunteers with 50 frames per cardiac cycle.
Synthesis Performance
The proposed model significantly outperformed traditional linear interpolation and optical flow methods. Magnitude image synthesis achieved a Peak Signal-to-Noise Ratio (PSNR) of 83.505, while phase (velocity) synthesis reached a PSNR of 63.727. Figure 2 illustrates the visual fidelity of the synthesized frames compared to ground truth across different time points, showing minimal pixel-wise error.
Segmentation and Velocity Correlation
The automated segmentation achieved a Dice coefficient of 0.959, indicating nearly perfect overlap with expert manual tracings. Furthermore, the correlation between predicted and true velocities (velocity coefficient) was 0.889, confirming the clinical utility of the synthesized data for motion analysis.
Discussion and Conclusion
By coupling segmentation with frame synthesis, the model leverages structural priors to improve motion estimation. While phase image synthesis remains challenging due to the lack of sharp structural edges compared to magnitude images, this multi-task approach provides a significant step forward. Future work will explore cross-modality synthesis to potentially generate velocity maps directly from standard magnitude images, further streamlining cardiac MRI protocols.
Summary updated