Generalizable, sequence-invariant deep learning image reconstruction for subspace-constrained quantitative MRI.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Deep Learning

Background:

• Magnetic Resonance (MR) Multitasking enables dynamic physiological mapping.
• Reconstructing images from diverse MR pulse sequences presents a challenge for deep learning models.
• Existing methods often struggle with cross-sequence generalizability and data scarcity.

Purpose of the Study:

• To develop a deep subspace learning network capable of functioning across various MR pulse sequences.
• To enhance the performance and generalizability of MR Multitasking image reconstruction.

Main Methods:

• A contrast-invariant component-by-component (CBC) network was developed and compared to a spatiotemporal multicomponent (MC) network.
• Experiments included matched-sequence (training/testing on same sequence) and unmatched-sequence (cross-sequence) evaluations.
• A universal CBC network was trained on mixed sequences (T1, T1-T2, T1-T2-fat fraction) for reconstruction.

Main Results:

• The CBC network demonstrated significantly superior performance (lower normalized root mean squared error) compared to the MC network across all experimental conditions.
• CBC exhibited enhanced generalizability, showing smaller performance degradation in unmatched-sequence tests.
• A single universal CBC network successfully reconstructed images from all tested pulse sequences, performing comparably to sequence-specific models.

Conclusions:

• Contrast-invariant learning of spatial features, rather than spatiotemporal features, improves MR Multitasking image reconstruction.
• The CBC approach enhances model performance, generalizability, and addresses data scarcity issues.
• This deep subspace learning framework offers a pathway towards universal, supervised reconstruction across diverse MR imaging sequences.