Skip to main content
2025 SCMR Conference Main banner
Final Program


Schedule at a Glance
Back

Rapid Fire Abstracts

Rapid Fire 10- Rapid, Efficient Imaging II

ECG-triggered, highly under sampled contrast-enhanced MRA of the thoracic aorta using deep learning reconstruction (RF_SA_493)

Saturday, February 1, 2025
10:50 AM – 11:00 AM East Coast USA Time
Location: Blue Room PreFunction
Background: Although contrast minimization for clinical CMR examinations is aimed at, first pass MRA still has a role in specific clinical questions. Due to the contrast agent (CA) dynamics, an acquisition within the CA bolus passage is advantageous and restricts the acquisition time to a few seconds. A tradeoff in spatial resolution is often inevitable and gating the acquisition for a sharp display of the ascending aorta is often sacrificed to achieve a feasible breath-hold time.

Methods: A triggered, contrast-enhanced MRA is presented with a Poisson-disc incoherent sampling pattern and a net acceleration factor of 8. A parasagittal slab covering the aorta was acquired with the sampling pattern reordered in a center-in-out fashion with a time to central k-space of 5 seconds The k-space acquisition is synchronized with the CA bolus passage using a test-bolus in a way that the central part of k-space is acquired during the first pass. The research sequence acquisition parameters were: FOV: 350x284x96 mm3, acquired/reconstructed spatial resolution: 1.4x1.4x2/0.7x0.7x1 mm3, scan window: 413 ms, flip angle: 17°, TE/TR=1.3/2.8ms. 12 datasets were acquired on a 3T MR scanner (MAGNETOM Vida, Siemens Healthineers, Forchheim, Germany) after informed consent in patients referred to a CMR examination. Online reconstruction was performed using a wavelet-regularized Compressed Sensing (CS) algorithm. Offline reconstruction was performed using an unrolled, deep learning-based (DL) algorithm. The latter comprises six iterations that alternate between data consistency based on a parallel imaging model and image regularization using three-dimensional U-nets. The network architecture was implemented in PyTorch and a supervised training performed using about 25000 training pairs derived from about 1000 fully sampled 3D datasets acquired from healthy volunteers on 1.5- and 3-T scanners (MAGNETOM scanners, Siemens Healthineers) in the head, abdomen, and pelvis. In alignment with data consistency principles, the network was tailored to enhance local image features, and as with clinically validated 2D methods, it was not expected to be sensitive to the content of the image. A blinded rating by two experts was performed on a 4-point Likert scale on the entire image (overall) as well as on a view focusing on the aortic outflow tract (aao, Fig. 2). The sharpness (1:non-diagnostic; 2:blurry; 3:sharp; 4:very sharp) and the artifact level (1:non-diagnostic; 2:strong artefacts; 3:low artefacts; 4:artefact-free) was rated. Additionally, for all datasets, the reader was asked to blindly pick the preferred volume CS vs. DL.

Results: All reconstructions were performed in under 30 seconds per dataset. Sagittal views and MIPs of two datasets are shown in Fig. 1 whereas Fig. 2 shows aortic outflow views comparing CS and DL. In the DL views, an improvement in image sharpness is clearly visible as well as a minimization of incoherent aliasing artefacts and motion artefacts as compared to the CS reconstructions. Fig. 3 summarises the rating results and demonstrates an improved rating score for the DL reconstructions as well as an overall improved image quality.

Conclusion: The presented DL reconstruction for first-pass CE-MRA shows significantly increased image quality as compared to the classical CS reconstruction. The reconstruction times being comparable, the DL reconstruction holds potential to replace CS for incoherently undersampled CE-MRA.