On the Foundation Model for Cardiac MRI Reconstruction

TL;DR

A foundation model that uses adaptive unrolling, channel-shifting, and Pattern and Contrast-Prompt-UNet (PCP-UNet) to tackle the problem of degraded image quality in cardiac magnetic resonance (CMR) imaging is proposed.

Abstract

In recent years, machine learning (ML) based reconstruction has been widely investigated and employed in cardiac magnetic resonance (CMR) imaging. ML-based reconstructions can deliver clinically acceptable image quality under substantially accelerated scans. ML-based reconstruction, however, also requires substantial data and computational time to train the neural network, which is often optimized for a fixed acceleration rate or image contrast. In practice, imaging parameters are often tuned to best suit the diagnosis, which may differ from the training data. This can result in degraded image quality, and multiple trained networks are needed to fulfill the clinical demands. In this study, we propose a foundation model that uses adaptive unrolling, channel-shifting, and Pattern and Contrast-Prompt-UNet (PCP-UNet) to tackle the problem. In particular, the undersampled data goes through a different number of unrolled iterations according to its acceleration rate. Channel-shifting improves reconstructed data quality. The PCP-UNet is equipped with an image contrast and sampling pattern prompt. In vivo CMR experiments were performed using mixed combinations of image contrasts, acceleration rates, and (under)sampling patterns. The proposed foundation model has significantly improved image quality for a wide range of CMR protocols and outperforms the conventional ML-based method.

Figures (6)

• Figure 1: Adaptive unrolling. Undersampled measurements are fed to a specific unrolled iteration according to the acceleration rate. Images of higher acceleration rate proceed through more unrolled iterations than those of lower acceleration rates. Each unrolled iteration has its own network regularizer and conjugate gradient with parameters independently determined relative to the other unrolled iterations.
• Figure 2: Pattern and Contrast-Prompt (PCP) UNet. PCP-UNet inherits the Prompt-UNet concept with additional k-space undersampling pattern prompt modules that are similar to the contrast prompt modules in Prompt-UNet.
• Figure 3: k-space sampling pattern embeddings were estimated from statistical information of the sampling mask, with an example of statistics along $k_x$ direction. The sampling mask is first split in two halves along the $k_x$ direction, in order to distinguish between uniform and random sampling. Summation along $k_x$ is applied to obtain the total number of samples as well as its distribution. The mean and variance of the sample number and the spacing between non-zero data points are computed as statistical feature of the sampling mask.
• Figure 4: Enlarging receptive field via channel-shifting. Circular shifting along the sub-sampled direction(s) is performed to produce shifted replicas of the input image. The replicas are subsequently concatenated with the input along channel dimension to produce an augmented input, which is fed into a regular CNN with additional channels in its input layer.
• Figure 5: Representative reconstructed image results for selected combinations of sampling patterns, acceleration rates, and contrasts, using all tested methods: FU: Fixed amount of UI with a UNet regularizer; AU: Adaptive amount of UI with a UNet regularizer; FP: Fixed amount of UI with the PCP-UNet regularizer; AP: Adaptive amount of UI with the PCP-UNet regularizer.