To use Convex Optimized Diffusion Encoding (CODE) to improve SNR in Diffusion-Weighted MRI (DWI) in the brain by eliminating sequence dead time and reducing TE.Neuro DWI applications typically uses a spin echo sequence with a single shot EPI readout (SE-EPI DWI) for short scans. High b-value imaging (b≥1000s/mm²) inherently reduces SNR, therefore it is important to minimize echo times (TE) to acquire high-quality images and Apparent Diffusion Coefficient (ADC) maps. SE-EPI DWI typically uses a diffusion encoding gradient waveform (G_Diff) with two identical components on either side of the 180° pulse. When combined with the temporal footprint of the EPI readout this leads to dead time between the 90° and 180° pulses. This dead time can be used for additional diffusion encoding thereby reducing TE for the same b-value and imaging parameters. Herein, a convex gradient optimized diffusion encoding (CODE) framework[1, 2] was used to design G_Diff that minimizes TE while conforming to the requirements of a SE-EPI DWI sequence.

Gradient Design: Convex optimization was used to to calculate G_Diff for which TE is minimized while conforming to hardware (G_Max=74mT/m and SR_Max=50T/m/s), pulse sequence (G_Diff=0 during RF activity and EPI readout), diffusion encoding (b-value) constraints. Volunteer Imaging: Healthy volunteers (N=10) were scanned on a 3.0T scanner (Siemens Prisma) after providing written informed consent for an IRB approved study. Single slice, axial DWI were acquired in the brain using b=1000s/mm² in three orthogonal directions, 1.6x1.6x3mm spatial resolution, 220x220mm FOV, BW=1450Hz/px, full-Fourier, GRAPPA acceleration factor 2, and TR=5000ms. DWI were acquired with both monopolar (MONO) (TE=75ms, Fig. 1A) and CODE (TE=64ms, Fig. 1B) diffusion encoding and were each repeated ten times for SNR analysis.

Reconstruction and Data Analysis: In vivo ADC maps were reconstructed individually for each of the ten repetitions of MONO and CODE. A voxel-wise SNR map was generated by dividing the mean ADC at each voxel by the standard deviation (across the ten repetitions). The global mean SNR was then calculated within the brain for CODE and MONO and compared across the ten subjects.

For b=1000s/mm² and 1.6x1.6x3mm spatial resolution, MONO required TE=75ms whereas CODE reduced this to TE=67ms (11% reduction, Fig 1). The TE reduction was 11% for b=2000s/mm² (80 vs 71ms) and 8% for b=5000s/mm² (89 vs 82ms) based on GDiff designs. In vivo, CODE produced b=0 images with 33±9% higher mean SNR than MONO (41.8±6.3 vs 31.3±4.4, P<0.001) ADC maps with 35±12% higher mean SNR than MONO (19.5±2.5 vs 14.5±1.9, P<0.0001) (Fig. 2) across the ten volunteers scanned.

CODE improved the SNR of ADC maps by reducing TE compared to MONO at the same b-value. The shortened TEs can potentially reduce the number of averages needed for acceptable SNR in DWI with large b-values. Alternatively, b-value could have been increased without significantly lengthening TE. Note that while only single slice imaging was performed, CODE is fully compatible with 2D multi-slice imaging and the shorter TE would permit more interleaved slices per TR.

A phantom study was also performed which showed excellent agreement between ADC values reported by MONO and CODE across a range of diffusivities (results not shown). Eddy current distortion artifacts were not apparent in CODE images when compared to MONO.

Gradient optimizations were performed for a clinical scanner with state-of-the-art gradient hardware (Siemens Prisma, Gmax=74mT/m). While longer TEs will result for conventional systems (Gmax=40mT/m), CODE still provides a similar benefit compared with MONO (e.g. TE=85ms vs 77ms for the b=1000s/mm² protocol described above).

All slew rates were constrained to ≤50T/m/s, which is significantly less than the 200T/m/s capability of the gradient system. This is a conservative bound that is software-imposed on all diffusion encoding gradients to avoid peripheral nerve stimulation (PNS). Our preliminary work shows that TEs can be further reduced if this constraint is relaxed, which can likely be done safely. Recent work has shown that a more sophisticated PNS model based on nerve response functions can be applied to safely shorten gradient waveforms[3]. Investigation of this approach in combination with the CODE framework is underway.

The CODE framework produces time-optimal diffusion encoding gradient waveforms that improve SNR by eliminating sequence dead times and shortening TEs.