Reconstruction of Under-Sampled Images and Concurrent Optimization of Sampling Masks for 3D Carotid Simultaneous...

To improve the imaging efficiency of 3D carotid simultaneous noncontrast angiography and intraplaque hemorrhage (SNAP) MRI by reconstruction of under-sampled images and concurrent optimization of sampling masks for the two shots of SNAP respectively.

A model-based deep learning architecture (deepSNAP) was proposed to recover under-sampled 3D carotid SNAP MRI. Sampling locations on the ky-kz plane were parameterized to enable respective optimization of the sampling masks for the two shots. A dataset of 100 3D carotid SNAP MRI scans was utilized (80 training, 20 test). Image recovery performance was compared with established techniques under different acceleration factors. Lumen area measurement accuracy and intraplaque hemorrhage (IPH) identification were evaluated at 6× acceleration. Prospective feasibility was assessed in 10 healthy volunteers with quantitative comparison against established methods.

deepSNAP exhibited superior image recovery performance on the test set, surpassing all comparison methods. Optimized masks generated by deepSNAP improved reconstruction performance across all comparison methods. High agreement between reconstructed images and original images was observed for lumen area measurement (ICC = 0.995, 95% CI: 0.993-0.996) and IPH detection (Cohen's κ = 0.976, 95% CI: 0.943-1.000). In the prospective experiment, deepSNAP achieved promising image quality and structural fidelity.

The deepSNAP model achieved under-sampled image reconstruction and simultaneous sampling mask optimization for SNAP, ensured the clinical practicability of the reconstructed images, and demonstrated preliminary technical feasibility in a prospective setting.