Skip to main content
SCMR 29th Annual Scientific Sessions Main banner
Final Program


Program-At-A-Glance

Oral Abstracts Session

Virtual Recording

Session: Oral Abstracts Science Session 5

Towards faster free-breathing multislice myocardial diffusion tensor imaging using deep-learning denoising reconstruction

Saturday, February 7, 2026
11:20 AM - 11:30 AM  
Location: Americas I and Hall 3 (2nd Lower Level)
Background: Diffusion tensor imaging (DTI) delivers unique insights into myocardial microstructure (1). Clinical applications need fast free breathing (FB) DTI scanning, which may be achieved by reducing the number of averages with deep-learning (DL) denoising reconstruction (2,3). We have set out to investigate the possible benefits and consequences in scatter and bias of using such an approach (ReconDL)(4-6) for the results of DTI analysis.

Methods: Free-breathing 3-slice second-order motion-compensated spin-echo (M2SE) DTI (7) of 5 healthy volunteers was acquired at 3T (GE SIGNA Premier, 8 x 12dir b=400s/mm2, 8 x 3dir b=50s/mm2, 2.3x2.3x8mm3 resolution, TE=82ms, end-systole, 360 cardiac cycles, at parallel-acceleration (ASSET) factor 3). This work re-used data originally acquired for testing GE DTI towards the multicentre SCMR SIGNET DTI study (8). The 8-averaged DTI acquisition raw-data was reconstructed with and without a deep-learning based denoising method (ReconDL, “DLOn” and “NoDL”, Fig 1). DTI analyses were separately repeated using the DLOn and NoDL magnitude images of all 8 averages, followed by further separate analyses for DLOn and NoDL at reduced numbers of averages for each, down to only 2 averages each. Each DTI analysis (by dedicated Matlab in-house software) (9) applied identical rejections of respiratory extremes, motion-corrections and ROIs in 6 SAX segments for each slice, total 90 ROIs. The results of each DTI analysis were compared to those from standard 8-average (NoDL) original acquisitions.

Results: Mean diffusivity (MD), fractional anisotropy (FA) and helix angle (HA) were as described in literature (1) for 8 averages without DL and the reduced 6 averages using ReconDL (Fig. 3). The ReconDL raw-data reconstruction method produced visible denoising of diffusion-weighted images (Fig. 1, left panels). Progressive reduction of the number of averages increased FA (10) and the measures of variability (Fig. 2). On Fig. 2, the boxes are derived from the 90 myocardial ROIs of each repeated DTI analysis, at the 1st and 3rd quartiles with median lines. The whiskers are each 1.5 x IQR and outliers are not shown. The application of ReconDL was effectively found to compensate for ≈25% reduction of acquisition time (i.e. down to 6 from 8 original repetitions, as highlighted by the curved arrows and horizontal grey bars on Fig. 2). Further acceleration may be possible with matched experimental settings.

Conclusion: We demonstrated feasibility of substantial reduction in DTI scan time using ReconDL. The DTI outputs were not biased by applying ReconDL, with the exception of FA due to altered SNR (10). Subject to confirmation in a larger cohort, we found that ReconDL enables approximately 25% reduction in free-breathing scanning time while preserving the normal values established in the fully sampled DTI scan.

Comparison of diffusion outputs showing equivalence of faster 6-averages (DL On) versus the reference 8 averages without ReconDL.























Normal subjects (n=5)MD µm2/msFAabs(E2A)
DL On6 averages1.54±0.170.41±0.1045±25°
No DL8 averages1.55±0.170.40±0.1046±27°
Fig.1 Typical example of the ReconDL raw-data reconstruction (upper half of Fig 1) reducing noise in the diffusion-weighted magnitude images (left greyscale panels) compared to reconstruction without ReconDL (lower left). The “DLOn” and “NoDL” reconstructions were sent into separate DTI analyses (right panels): mean diffusivity (MD), fractional anisotropy (FA) and helical angle (HA) maps, and DTI analyses repeated with reduced averaging, shown here for 8 and 2 averages.
Fig.2 The impacts of reduced averaging (i.e. shorter scanning time) combined with ReconDL. The grey horizontal bands highlight the standard ranges of the 8 average NoDL scan. Curved arrows indicate the observed ReconDL’s potential to reduce scan time by approximately 25% while preserving the DTI measurement properties.
Fig.3 Comparison of diffusion outputs showing equivalence of 6-averages (DL On) versus the reference 8 averages without ReconDL.