DCRA-Net: Attention-Enabled Reconstruction Model for Dynamic Fetal Cardiac MRI

TL;DR

DCRA-Net addresses the challenge of reconstructing dynamic fetal cardiac motion from highly accelerated, non-gated MRI by integrating spatial and temporal self-attention with temporal-frequency data representations. The model outperforms conventional and DL baselines (L+S, k-GIN, and 3D U-Net) across fetal and adult datasets, achieving peak SSIM values near 0.989 and 0.964 under lattice undersampling for fetal and adult data, respectively. An ablation study shows that temporal-frequency representation plus data consistency synergistically improve reconstruction quality, and the method demonstrates potential for broader dynamic MRI applications beyond fetal imaging. The work also highlights the importance of acquisition pattern and identifies remaining challenges in generalization to diverse undersampling schemes and multi-coil data, guiding future clinical translation.

Abstract

Dynamic fetal heart magnetic resonance imaging (MRI) presents unique challenges due to the fast heart rate of the fetus compared to adult subjects and uncontrolled fetal motion. This requires high temporal and spatial resolutions over a large field of view, in order to encompass surrounding maternal anatomy. In this work, we introduce Dynamic Cardiac Reconstruction Attention Network (DCRA-Net) - a novel deep learning model that employs attention mechanisms in spatial and temporal domains and temporal frequency representation of data to reconstruct the dynamics of the fetal heart from highly accelerated free-running (non-gated) MRI acquisitions. DCRA-Net was trained on retrospectively undersampled complex-valued cardiac MRIs from 42 fetal subjects and separately from 153 adult subjects, and evaluated on data from 14 fetal and 39 adult subjects respectively. Its performance was compared to L+S and k-GIN methods in both fetal and adult cases for an undersampling factor of 8x. The proposed network performed better than the comparators for both fetal and adult data, for both regular lattice and centrally weighted random undersampling. Aliased signals due to the undersampling were comprehensively resolved, and both the spatial details of the heart and its temporal dynamics were recovered with high fidelity. The highest performance was achieved when using lattice undersampling, data consistency and temporal frequency representation, yielding PSNR of 38 for fetal and 35 for adult cases. Our method is publicly available at https://github.com/denproc/DCRA-Net.

Figures (5)

• Figure 1: The proposed DCRA-Net model for dynamic fetal cardiac MRI reconstruction. The main parts of the model are encoder, bottleneck and decoder parts. Encoder and decoder blocks are built from two ResNet blocks, spatial (S) self-attention block followed by temporal (T) self-attention layer and down-/upsampling. The bottleneck consists of ResNet block, spatial and temporal self-attention layers followed by another ResNet block. The initial block uses spatial convolutional layer and temporal self-attention block. Our implementation assumes number of channels $C = 64$ for both fetal and adult application.
• Figure 2: Examples of $8$x undersampling patterns.
• Figure 3: Fetal heart reconstruction comparison shows reconstructed data as image frames (a), temporal (c) and frequency (e) representations and their corresponding error maps (b, d, f). (e) and (f) are also shown using reduced dynamic range in (g, h).
• Figure 4: Adult heart reconstruction comparison shows reconstructed data as image frames (a), temporal (c) and frequency (e) representations and their corresponding error maps (b, d, f). (e) and (f) are also shown using reduced dynamic range in (g, h).
• Figure 5: Comparison of DCRA-Net across temporal representations and data consistency presence. The figure shows reconstructed data as image frames (a), temporal (c) and frequency (e) representations and their corresponding error maps (b, d, f). (e) and (f) are also shown using reduced dynamic range in (g, h).