Deep-Learning-Based Image Reconstruction to Improve End-Diastolic and Systolic Cardiac T1 Mapping

To develop an image reconstruction method that enables increased spatial resolution cardiac T1 mapping in both the end-diastolic and systolic phase, that shows high T1 agreement with the clinical standard. The resolution gain is achieved by increasing the acceleration rate of MOLLI single-shot images to R = 4, while maintaining a sufficiently short acquisition window.

A modified end-to-end variational network (MappingVN) is proposed. The modifications include a re-ordered sheared-grid sampling pattern, 2D + contrast convolutions and the use of patchwise squeeze-and-excitation layers. The method was evaluated in terms of image quality and T1 agreement with reference MOLLI T1 maps using retrospectively undersampled patient data. Furthermore, the method was additionally evaluated in a prospective setting comparing high-resolution T1 maps (1.14 × 1.14 mm²) to reference T1 maps in standard resolution (1.41 × 2.13 mm²). Finally, the applicability for systolic T1 mapping was explored using increased acceleration to shorten the acquisition window.

The MappingVN showed improved SSIM scores of 0.95 and 0.98 on 1.5T and 3 T compared to 0.93 and 0.96 for GRAPPA. In high-resolution end-diastolic T1 maps stronger T1 agreement (MappingVN: -3 ± 69 ms on 1.5T, -11 ± 70 ms on 3T, GRPPA: -9 ± 129 ms on 1.5T, 14 ± 106 ms) could be observed. For systolic T1 mapping the MappingVN reduced the occurrence of motion artifacts.

The proposed method enables high spatial resolution cardiac T1 mapping in both end-diastolic and systolic phases. Resulting maps show good T1 agreement with the clinical standard and may improve the visibility of small focal lesions while reducing partial volume effects.