1344769 - Universal sequence-invariant deep learning for CMR Multitasking image reconstruction

CMR Multitasking¹ is a promising technique for quantitative CMR without breath-holds or ECG. Practical reconstruction time is feasible with supervised deep subspace learning^2,3, but requires sequence-specific training. In this work, we explore whether universal, sequence-invariant CMR Multitasking deep learning reconstruction is possible by trading temporal awareness (breadth) for added depth. We evaluated the performance and generalizability of both strategies by training on T1 mapping data only, then testing on: a) matched-sequence T1 mapping data; and b) T1-T2 mapping data from a different sequence.

Methods:

Network structures: We evaluated the previous spatiotemporal “multi-component” (MC) network³ as well as spatial-only “component-by-component” (CBC) networks. Specifically, we evaluated three different U-Nets (Fig. 1):
1. MC3 – spatiotemporal, 3 downsampling (DS) steps, 1.1M weights
2. CBC3 – spatial, 3 DS steps, 3k weights
3. CBC8 – spatial, 8 DS steps, 1.1M weights

Training data: All networks were trained on Multitasking T1 mapping data collected from three Siemens 3T scanners: 118 cases for training and 13 cases for validation³.

Network comparison: First, we evaluated the impact of removing temporal awareness (MC3 to CBC3) and the impact of trading temporal awareness for depth (MC3 to CBC8) in 15 testing data collected using the matching T1-only pulse sequence. We then evaluated the generalizability of MC3 and CBC8 to a different pulse sequence by testing the T1-trained networks on 20 T1-T2 Multitasking data.

Evaluation methods: Image NRMSE was calculated using iteratively reconstructed images as a reference. Performance between networks was compared using the Wilcoxon signed-rank test, and performance between datasets was compared using the Mann–Whitney U test, both regarding p< 0.05 as significant. In the T1-T2 testing set, Bland-Altman analyses evaluated accuracy and precision of myocardial T1 and T2 values at end-diastole vs. corresponding T1 and T2 values from iterative reconstruction.

Results:

Fig. 2 shows network performance on both T1 testing data and T1-T2 testing data. On T1 data, CBC3 performed worse than MC3, but CBC8 performance was not significantly different from MC3 (Fig. 2A), indicating temporal awareness can be traded for depth. On T1-T2 data, CBC8 performed better than MC3 (Fig. 2B). MC3 performed worse on T1-T2 data than on T1 data, but CBC8 performance was not significantly different between testing sets (Fig. 2C). On T1-T2 maps, CBC8 produced fewer structural features in error maps and tighter limits of agreement by 29% in T1 and 32% in T2 vs. MC3 (Fig. 3).

Conclusion:

In our proposed deep learning reconstruction method, trading temporal awareness for network depth maintained reconstruction quality in T1 CMR Multitasking data, and allowed generalization to novel T1-T2 data without retraining. If the generalizability of CBC networks extends to multiple sites, vendors, and field strengths, it could facilitate clinical translation.