Generalizable, sequence-invariant deep learning image reconstruction for subspace-constrained quantitative MRI.
Authors
Affiliations (3)
Affiliations (3)
- Department of Radiological Sciences, David Geffen School of Medicine at UCLA, Los Angeles, California, USA.
- Department of Bioengineering, University of California, Los Angeles, California, USA.
- Biomedical Imaging Research Institute, Cedars-Sinai Medical Center, Los Angeles, California, USA.
Abstract
To develop a deep subspace learning network that can function across different pulse sequences. A contrast-invariant component-by-component (CBC) network structure was developed and compared against previously reported spatiotemporal multicomponent (MC) structure for reconstructing MR Multitasking images. A total of 130, 167, and 16 subjects were imaged using T<sub>1</sub>, T<sub>1</sub>-T<sub>2</sub>, and T<sub>1</sub>-T<sub>2</sub>- <math xmlns="http://www.w3.org/1998/Math/MathML"> <semantics> <mrow><msubsup><mi>T</mi> <mn>2</mn> <mo>*</mo></msubsup> </mrow> <annotation>$$ {\mathrm{T}}_2^{\ast } $$</annotation></semantics> </math> -fat fraction (FF) mapping sequences, respectively. We compared CBC and MC networks in matched-sequence experiments (same sequence for training and testing), then examined their cross-sequence performance and generalizability by unmatched-sequence experiments (different sequences for training and testing). A "universal" CBC network was also evaluated using mixed-sequence training (combining data from all three sequences). Evaluation metrics included image normalized root mean squared error and Bland-Altman analyses of end-diastolic maps, both versus iteratively reconstructed references. The proposed CBC showed significantly better normalized root mean squared error than MC in both matched-sequence and unmatched-sequence experiments (p < 0.001), fewer structural details in quantitative error maps, and tighter limits of agreement. CBC was more generalizable than MC (smaller performance loss; p = 0.006 in T<sub>1</sub> and p < 0.001 in T<sub>1</sub>-T<sub>2</sub> from matched-sequence testing to unmatched-sequence testing) and additionally allowed training of a single universal network to reconstruct images from any of the three pulse sequences. The mixed-sequence CBC network performed similarly to matched-sequence CBC in T<sub>1</sub> (p = 0.178) and T<sub>1</sub>-T<sub>2</sub> (p = 0121), where training data were plentiful, and performed better in T<sub>1</sub>-T<sub>2</sub>- <math xmlns="http://www.w3.org/1998/Math/MathML"> <semantics> <mrow><msubsup><mi>T</mi> <mn>2</mn> <mo>*</mo></msubsup> </mrow> <annotation>$$ {\mathrm{T}}_2^{\ast } $$</annotation></semantics> </math> -FF (p < 0.001) where training data were scarce. Contrast-invariant learning of spatial features rather than spatiotemporal features improves performance and generalizability, addresses data scarcity, and offers a pathway to universal supervised deep subspace learning.