Multi-dynamic deep image prior for cardiac MRI

Abstract

Purpose: Cardiovascular magnetic resonance imaging is a powerful diagnostic tool for assessing cardiac structure and function. However, traditional breath-held imaging protocols pose challenges for patients with arrhythmias or limited breath-holding capacity. This work aims to overcome these limitations by developing a reconstruction framework that enables high-quality imaging in free-breathing conditions for various dynamic cardiac MRI protocols.

Methods: Multi-Dynamic Deep Image Prior (M-DIP), a novel unsupervised reconstruction framework for accelerated real-time cardiac MRI, is introduced. To capture contrast or content variation, M-DIP first employs a spatial dictionary to synthesize a time-dependent intermediate image. Then, this intermediate image is further refined using time-dependent deformation fields that model cardiac and respiratory motion. Unlike prior DIP-based methods, M-DIP simultaneously captures physiological motion and frame-to-frame content variations, making it applicable to a wide range of dynamic applications.

Results: We validate M-DIP using simulated MRXCAT cine phantom data as well as free-breathing real-time cine, single-shot late gadolinium enhancement (LGE), and first-pass perfusion data from clinical patients. Comparative analyses against state-of-the-art supervised and unsupervised approaches demonstrate M-DIP's performance and versatility. M-DIP achieved better image quality metrics on phantom data, higher reader scores on in-vivo cine and LGE data, and comparable scores on in-vivo perfusion data relative to another DIP-based approach.

Conclusion: M-DIP enables high-quality reconstructions of real-time free-breathing cardiac MRI without requiring external training data. Its ability to model physiological motion and content variations makes it a promising approach for various dynamic imaging applications.

Figure 1:

Overview of M-DIP. (A) Trainable dynamic code vectors of dimensionality $K$, (B) trainable static code vectors of dimensionality $N\times c$, (C) the network ${𝒢}_{\psi }$ that generates a different deformation field for each $t$, (D) the network ${𝒢}_{\mathit{\zeta }}$ that generates a different set of weights for each $t$, (E) the network ${𝒢}_{\mathit{\theta }}$ that generates $L$ time-invariant spatial dictionary elements, (F) x- and y-components of the deformation field generated by ${𝒢}_{\psi }$ at $t=\tau$, (G) weights generated by ${𝒢}_{\zeta }$ at $t=\tau ,$ (H) spatial dictionary ${\mathbf{s}}^{(1:L)}$ generated by ${𝒢}_{\mathit{\theta }}$, (I) linear combination of the dictionary elements with time-specific weights at $t=\tau$, (J) intermediate image at $t=\tau$, (K) warping operation that applies deformation to the intermediate image, and (L) the output frame at $t=\tau$.

Figure 2:

M-DIP and LR-DIP results of the phantom study with results of Student’s t-tests. * indicates $p\le 0.05$, ** indicates $p\le 0.01$,*** indicates $p\le 0.001$, and **** indicates $p\le 0.0001$.

Figure 3:

Exemplary MRXCAT phantom reconstructions. Each sub-figure illustrates an end-diastolic frame, temporal profiles, and a close-up of the heart. Red arrows highlight an artifact present in LR-DIP that is suppressed in M-DIP.

Figure 4:

Bar plots with the scoring results for the in-vivo studies.

Figure 5:

Exemplary in-vivo real-time cine reconstructions. Each sub-figure illustrates an end-diastolic frame, temporal profiles, and a close-up of the heart. Red arrows show an area where differences in image sharpness are particularly apparent.

Figure 6:

Exemplary in-vivo free-breathing single-shot LGE reconstructions. Each sub-figure illustrates one frame, temporal profiles, and a close-up of the heart. Red arrows show an enhanced area that is better visible in M-DIP compared to LR-DIP.

Figure 7:

Exemplary in-vivo first-pass perfusion reconstructions. Each sub-figure illustrates one frame, temporal profiles, and a close-up of the heart. Red arrows show a septal perfusion defect clearly visible in all three reconstructions.

Figure 8:

M-DIP reconstructions of a real-time cine series with the deformation field generator ${𝒢}_{\psi }$ (a) activated and (b) deactivated, and M-DIP reconstructions of a perfusion dataset with spatial dictionaries of (c) size $L=24$ and (d) size $L=1$. The red arrows point to the artifact observed in the perfusion case when $L=1$.