Distributional Deep Learning for Super-Resolution of 4D Flow MRI under Domain Shift

TL;DR

The paper tackles the domain-shift problem in super-resolving 4D Flow MRI by introducing Distributional Super-Resolution (DSR), which expands the training domain through multi-covariate pre-additive noise and leverages a distributional loss to improve extrapolation. It formalizes Y = h(X + ε) with a semiparametric form h(t) = g(β^T t), proves distributional extrapolability and consistency, and extends to multivariate outcomes with a deep learning framework. The authors implement a patch-based, geometry-agnostic SR pipeline with self-supervised pre-training on CFD data, followed by two-step LP-FT fine-tuning on paired CFD-4DF data, and demonstrate superior performance over standard regression and 4DFlowNet in both simulations and real 4DF data. The approach yields robust SR under domain shift and offers practical utility for clinical 4DF analysis, with code and data resources provided for replication and broader adoption.

Abstract

Super-resolution is widely used in medical imaging to enhance low-quality data, reducing scan time and improving abnormality detection. Conventional super-resolution approaches typically rely on paired datasets of downsampled and original high resolution images, training models to reconstruct high resolution images from their artificially degraded counterparts. However, in real-world clinical settings, low resolution data often arise from acquisition mechanisms that differ significantly from simple downsampling. As a result, these inputs may lie outside the domain of the training data, leading to poor model generalization due to domain shift. To address this limitation, we propose a distributional deep learning framework that improves model robustness and domain generalization. We develop this approch for enhancing the resolution of 4D Flow MRI (4DF). This is a novel imaging modality that captures hemodynamic flow velocity and clinically relevant metrics such as vessel wall stress. These metrics are critical for assessing aneurysm rupture risk. Our model is initially trained on high resolution computational fluid dynamics (CFD) simulations and their downsampled counterparts. It is then fine-tuned on a small, harmonized dataset of paired 4D Flow MRI and CFD samples. We derive the theoretical properties of our distributional estimators and demonstrate that our framework significantly outperforms traditional deep learning approaches through real data applications. This highlights the effectiveness of distributional learning in addressing domain shift and improving super-resolution performance in clinically realistic scenarios.

Figures (8)

• Figure 1: Flow chart for distributional super-resolution procedure. Data Preparation: CFD training pairs are generated through cropping, data augmentation and downsampling. Pre-training: These pairs are used to pre-train the proposed DSR model. Fine-tuning: A small subset of paired 4DF and CFD patches is used to refine the model's estimators. Evaluation: The fine-tuned model is then evaluated on the complete 4DF testing dataset.
• Figure 2: The 3D U-Net architecture. The convolutional layer in the red boxes is replaced by new layer in the fine-tuning process.
• Figure 3: The comparision between DSR and $L_2$ regression models under the low dimensional setting. The gray dots denote the training data, red line denotes the prediction with true parameters, blue line denotes the prediction with estimated parameters from DSR or $L_2$ regression models and the shaded areas denote the 10% to 90% quantile interval.
• Figure 4: The comparision between DSR and $L_2$ regression models under the high dimensional setting. The gray dots denote the training data, red line denotes the prediction with true parameters, blue line denotes the prediction with estimated parameters from DSR or $L_2$ regression models and the shaded areas denote the 10% to 90% quantile interval.
• Figure 5: The performance of DSR and $L_2$ regression model on the validation data. The magnitude is defined as the vector norm of the values along the three spatial directions. (a) Comparison of fitted density estimates between the DSR model and the $L_2$ model. (b) MSE comparison between the prediction error of DSR model and the $L_2$ model. (c) Absolute error heatmaps for a representative subject. Blue box regions in which DSR achieves lower errors than the $L_2$ model baseline.

Theorems & Definitions (4)

• Definition 3.1: Distributional extrapolability
• Theorem 1: Distributional extrapolability of DSR
• Theorem 2: Estimation Consistency
• Remark