Self-Supervised Adversarial Diffusion Models for Fast MRI Reconstruction

TL;DR

This work tackles the challenge of speeding up MRI acquisition without depending on fully sampled training data. It introduces SSAD-MRI, a self-supervised adversarial diffusion framework that employs an adversarial mapper within a diffusion model and a self-supervised training scheme using adaptive sampling masks. The method achieves superior reconstruction quality across within-domain and out-of-domain brain MRI datasets, particularly at higher acceleration factors, with statistically significant gains in NMSE, PSNR, and SSIM. The findings suggest meaningful clinical implications for faster, higher-quality MRI, though prospective 3D validation and deployment considerations remain for future work.

Abstract

Purpose: To propose a self-supervised deep learning-based compressed sensing MRI (DL-based CS-MRI) method named "Adaptive Self-Supervised Consistency Guided Diffusion Model (ASSCGD)" to accelerate data acquisition without requiring fully sampled datasets. Materials and Methods: We used the fastMRI multi-coil brain axial T2-weighted (T2-w) dataset from 1,376 cases and single-coil brain quantitative magnetization prepared 2 rapid acquisition gradient echoes (MP2RAGE) T1 maps from 318 cases to train and test our model. Robustness against domain shift was evaluated using two out-of-distribution (OOD) datasets: multi-coil brain axial postcontrast T1 -weighted (T1c) dataset from 50 cases and axial T1-weighted (T1-w) dataset from 50 patients. Data were retrospectively subsampled at acceleration rates R in {2x, 4x, 8x}. ASSCGD partitions a random sampling pattern into two disjoint sets, ensuring data consistency during training. We compared our method with ReconFormer Transformer and SS-MRI, assessing performance using normalized mean squared error (NMSE), peak signal-to-noise ratio (PSNR), and structural similarity index (SSIM). Statistical tests included one-way analysis of variance (ANOVA) and multi-comparison Tukey's Honesty Significant Difference (HSD) tests. Results: ASSCGD preserved fine structures and brain abnormalities visually better than comparative methods at R = 8x for both multi-coil and single-coil datasets. It achieved the lowest NMSE at R in {4x, 8x}, and the highest PSNR and SSIM values at all acceleration rates for the multi-coil dataset. Similar trends were observed for the single-coil dataset, though SSIM values were comparable to ReconFormer at R in {2x, 8x}. These results were further confirmed by the voxel-wise correlation scatter plots. OOD results showed significant (p << 10^-5 ) improvements in undersampled image quality after reconstruction.

Figures (4)

• Figure 1: The flowchart of our proposed self-supervised adaptive diffusion model is illustrated. The coil sensitivity $\mathcal{S}$, random sampling mask $\Omega$, and k-space were illustrated as inputs. Firstly, the random sampling pattern $\Omega$ is divided into two non-overlapping sampling masks $\aleph$ and $\Upsilon$ that were used in the training path and loss path as given in (\ref{['eq:kl_loss_03']}), respectively. Then, the adversarial mapper was trained as given in (\ref{['eq:discloss_01']}) and (\ref{['eq:loss_gen_01']}) using the data sampled in step $t+k$. The sample $y_{t+k}$ used to recover $\hat{y}_0$ where was used to calculate $\hat{y}_t$ in a given step $t$ Equation (\ref{['eq:sampling_qt_from_q0']}).
• Figure 2: Within-domain axial T2-w image reconstruction at $R=2\times$, $4\times$, and $8\times$ and $\rho=0.5$ are illustrated. (a) illustrates the results for a healthy subject and (b) illustrates the difference map between the reconstructed and ground truth images.
• Figure 3: Reconstruction of within-domain MP2RAGE T1 quantitative maps at acceleration factors $R \in \{ 2\times, 4\times, 8\times\}$ with $\rho=0.5$. (a) Reconstructed images for a representative subject. (b) Corresponding pixel-wise difference maps between the reconstructed images and the ground truth images.
• Figure 4: Out-of-distribution reconstruction quantitative results are illustrated for multi-coil axial T1-w and T1c in the first and second rows, respectively, for three difference acceleration rates. The red stars indicate statistically significant differences (p$< 0.05$).