Convolutional Network Based Lineshape Adaptation Improved Subject-Specific Subspace Estimation for MRSI Reconstruction

Motivation: In subspace-based MRSI, there is potential discrepancy between a physics-driven pre-learned subspace for reconstruction and experimental variations for individual experiments. Goal(s): To develop a strategy to accurately adapt the pre-learned subspace to subject/experiment-dependent spectral variations while preserving weak metabolite signals and avoiding overfitting. Approach: A CNN-based strategy was proposed, incorporating spatial constraints and a union-of-subspace-based signal separation/protection, to achieve effective subspace adaptation for individual experiments. Results: Our method achieved lower adaptation errors, better preservation of weak metabolite signals and improved quantification. Impact: The proposed method minimized the discrepancy between physics-driven subspace and in vivo data, improving subspace-based MRSI reconstruction and quantification.