VTransformer: A Universal Field- and Age-Agnostic Foundation Model for 5T and Pediatric Cardiac MRI Reconstruction

Background: The CMRxRecon 2025 challenge aims to develop generalizable foundational image reconstruction models across diverse pulse sequences, centers, field strengths, vendors, and diseases. The two special tasks are: (a) 5T cardiac MRI reconstruction from 1.5T/3T training data, and (b) pediatric image reconstruction from adult training data. We propose VTransformer, a vision transformer-based foundation model, and compare its performance against the CMRxRecon 2024 winner (PromptMR-plus) and parallel imaging methods (GRAPPA and SENSE).

Methods: We developed a deep unrolled cardiac MRI reconstruction backbone with Vision Transformer–Channel Attention blocks (Figure 1). To handle domain shifts, we combine adaptive normalization for field-dependent variations, multi-scale feature extraction for motion and size differences, temporal-consistency mechanisms for varying heart rates, and multi-contrast attention for chamber proportion changes. The backbone trains exclusively on standard 1.5T/3T adult data without 5T and pediatric exposure, with validation following held-out protocols for cross-domain assessment.

Results: On the 5T validation set (n=241), our model achieved SSIM 0.804 ± 0.107, PSNR 29.63 ± 4.54 dB, NMSE 0.043 ± 0.038, significantly outperforming all baselines (paired t-test): PromptMR-plus (SSIM 0.607 ± 0.255, PSNR 24.53 ± 7.31 dB, NMSE 0.753 ± 0.927; p< 0.001), SENSE (SSIM 0.686 ± 0.150, PSNR 28.02 ± 4.46 dB, NMSE 0.102 ± 0.053; p< 0.001), GRAPPA (SSIM 0.750 ± 0.102, PSNR 28.80 ± 4.09 dB, NMSE 0.067 ± 0.029; p< 0.01), and zero-filled (SSIM 0.661 ± 0.097, PSNR 26.15 ± 3.22 dB, NMSE 0.312 ± 0.055; p< 0.001). PSNR improvements ranged from 0.83-5.10 dB with substantial NMSE reductions across all methods.

On the pediatric validation set (n=542), the model reached SSIM 0.787 ± 0.117, PSNR 27.75 ± 4.41 dB, NMSE 0.073 ± 0.089, with highly significant improvements over all baselines (all p< 0.001): PromptMR-plus (SSIM 0.431 ± 0.122, PSNR 16.10 ± 2.60 dB, NMSE 1.228 ± 0.937), SENSE (SSIM 0.341 ± 0.143, PSNR 15.96 ± 2.48 dB, NMSE 1.156 ± 0.197), GRAPPA (SSIM 0.389 ± 0.135, PSNR 18.12 ± 2.68 dB, NMSE 0.945 ± 0.199), and zero-filled (SSIM 0.378 ± 0.125, PSNR 17.23 ± 2.52 dB, NMSE 1.087 ± 0.200). PSNR gains of 9.63-11.79 dB demonstrated large practical improvements. All comparisons showed large effect sizes and statistical significance was robust across different measurement scenarios (Figure 2). Representative visual comparisons demonstrate the superior reconstruction quality of our approach across both challenging domains (Figure 3).

Conclusion: We present a universal foundation model for cardiac MRI reconstruction achieving exceptional cross-domain generalization without domain-specific training. Our transformer-enhanced architecture, trained only on standard 1.5T/3T adult data, delivers 32% SSIM improvement at 5T and 82% improvement in pediatric cases versus state-of-the-art methods.




Figure 1: Architecture overview. (A) The overall framework processes nine adjacent cardiac MRI contrasts (T1w, T2w, Cine, LGE, etc.) through N cascaded reconstruction blocks, each incorporating Adaptive Contrast Sampling (ACS), Sensitivity Map Estimation (SME), inverse Fourier transform (F⁻¹), and Root-Sum-of-Squares (RSS) combination to produce the final reconstruction. (B) Individual cascade structure showing dual-path processing with high-resolution feature extraction for spatial detail and low-resolution feature extraction for global context, followed by an Advanced Denoiser and data consistency (DC) enforcement with learned step size (η'). (C) Detailed Advanced Denoiser architecture implementing PromptUnet with Vision Transformer-Channel Attention Blocks (ViT-CAB), featuring DownBlocks, UpBlocks, PromptBlocks, and BottleneckBlocks with momentum connections and multi-scale feature processing for adaptive fusion of global and local representations.
Figure 2: Performance comparison across reconstruction methods for specialized cardiac MRI applications. (A) Results for 5T high-field cardiac MRI reconstruction showing PSNR (left), SSIM (middle), and NMSE (right) metrics. (B) Results for pediatric cardiac MRI reconstruction with the same metric layout. Methods compared include Zero-Filled (ZF), SENSE, GRAPPA, PromptMR-Plus, and our proposed VTransFormer approach (Ours).
Figure 3: Qualitative comparison of cardiac MRI reconstruction methods across specialized applications. (A) Visual results for 5T high-field cardiac MRI reconstruction comparing Zero-Filled (ZF), SENSE, GRAPPA, PromptMR-plus, and our approach (Ours) across representative samples from TaskS1. (B) Visual results for pediatric cardiac MRI reconstruction using the same methods across representative samples from TaskS2. Each reconstructed image displays corresponding SSIM and PSNR values. Our approach consistently produces the clearest reconstructions with superior preservation of anatomical details and minimal artifacts across both challenging domains. Traditional methods (SENSE, GRAPPA) exhibit significant noise and blurring, while PromptMR-plus shows inconsistent performance with incomplete reconstruction in several cases. Our method demonstrates robust generalization capabilities, achieving the highest quantitative scores while maintaining excellent visual quality across diverse cardiac anatomies and imaging conditions without domain-specific training.