Quick Fire Session
SCMR 22nd Annual Scientific Sessions
Novel MRI approaches using Deep Neural Networks1-3 (DNN) typically assume: 1) high-fidelity of training data (i.e. fully sampled, artifact-free images), and 2) optimized DNN models. For CMR, clinical imaging constraints, including subject motion and systems hardware imperfections may yield suboptimal training data. We propose a new reconstruction approach using DNN that: a) accounts for imperfect CMR training data fidelity, and b) does not require a carefully tuned DNN model. Our approach deliberately exploits training-derived remnant artifacts as vectorized projections that are separable from signal source. Using non-Cartesian high-temporal Cine-CMR4,5 (~15ms/phase) example, we demonstrate one application of our generalizable framework that may mitigate conditions for training data fidelity and need for elaborate DNN modeling.
In brief, our generalized approach incorporates DNN-based training that deliberately projects unresolved reconstruction artifacts onto known artifact-vectors under a mathematical transform domain; we thereby separate orthogonally aligned artifact-vectors from the invariant signal source. For this Radial cine-CMR example (n=10 subjects; 14 slices x 80 cardiac phases; 9-training+1-validation cycled for every subject), we employed for DNN training our Nyquist-satisfied reconstructions that employed a weighted [KWIC] temporal window6 that nonetheless contained unresolved reconstruction artifacts. R=5 accelerated re-gridded data was used for DNN input. Schematic of artifact vector projection example is shown in Figure 1, derived via DNN in the image x- and y-directions for illustrative clarity. These artifact vectors can be yielded in two orthogonal directions, and filtered from the invariant source (e.g. LPF, blind-source separation [PCA, ICA], etc.). For the proposed approach assessment, invariance of Gaussian noise distribution was first examined in every output frame using the Kolmogorov-Smirnov (KS) test7. A per-cardiac phase apparent Signal-to-Noise-Ratio (aSNR) map was next generated to yield a fair map ratio for comparisons between alternative reconstructions4.
Figure 2a shows the KS-test map that satisfied the Gaussian noise condition for the subsequent SNR-map-based assessment in a fair and controlled manner across alternative reconstruction techniques. Figure 2b shows: a) input, b) training, c) previous4, and d) proposed methods. The derived aSNR-map ratios across 10x14x80 frames were: 10.33 (proposed/training); 5.33 (proposed/previous). Both comparisons were statistically significant (p<0.0005).
We demonstrate approach that exploits signal invariance and artifact vectorization for CMR using imperfectly trained DNNs. This example yielded high-frame-rate cine-CMR reconstructions with superior radial streaking artifact suppression compared to previous methods, despite the presence of unresolved radial streaking in the employed training dataset as answer key.