Skip to content

About

Ultra accelerated single shot 3D EPI sequences, plus an accompanying multi-echo GRE sequence for Xiaokai's COMET project.

Resources

Stars

0 stars

Watchers

0 watching

Forks

Repository files navigation

2026-11-12 comet session — sequence summary

Sequences for the three items the collaborator requested, built 2026-11-11 and revised 2026-11-12 with ArbEPI-python main (28a28fa, unmodified; fetched and up to date) by generate.py in this folder. Protocol for every EPI: 90 x 90 x 60 matrix, 2.4 mm isotropic, FOV 216 x 216 x 144 mm, GE_MR750, water excitation (no fat-sat), human PNS weights, no discard frames, 5 min (Nframes = ceil(300 s / frame TR)), 4 µs ADC dwell for all three. Every file is under the 80 % normal-mode PNS limit (generate.py refuses to export above it). The .pge files were copied to the inside scanner with ge/coppe.py (entry numbers 266-275, see "Scanner transfer").

Folder Request Trajectory Sampling ETL x Nshots R TE of k = 0 Frame TR Nframes (scan) Flip Dwell Peak PNS ArbEPI / EPIcal Acoustics ArbEPI / EPIcal (limit 0.3)
1_laminar_1x_2.4mm_ETL90 1 laminar, one ky line per train fully sampled 90 x 60 1 33.301 ms 62.830 ms/shot, 3.7697 s volume 80 (301.6 s) 17.67° 4 µs 69.6 % / 69.6 % 0.090 / 0.089
2_radial_90x_2.4mm_ETL60 2 radial, one train per frame mask_60, 60 samples 60 x 1 90.0 30.002 ms 54.744 ms 5481 (300.1 s) 16.51° 4 µs 76.5 % / 74.7 % 0.039 / 0.040
3_radial_122.7x_2.4mm_ETL44 2 radial, one train per frame mask_44, 44 samples 44 x 1 122.7 30.000 ms 48.844 ms 6143 (300.0 s) 15.61° 4 µs 76.7 % / 73.3 % 0.044 / 0.045
4_meGRE_dwell4us 3 monopolar multi-echo GRE Cartesian 90 x 90 x 60 22 echoes, 1 excitation/(ky,kz) -- TE 2.236 ... 49.192 ms 51.35 ms 291.2 s 16.0° 4 µs 71.6 % 0.155

Folders 1-3 hold ArbEPI, EPIcal, noise (.seq + .pge), scan_info.mat, the diagnostic plots (mask.png, trajectory.png, psf.png, one_tr.png, PNS_one_tr.png) and build_info.json (every timing number). Folder 4 holds meGRE.seq/.pge, meGRE_info.json and one_tr.png (the first image excitation: prephasers, 22 echoes with their centers marked, spoiler/rewinders). logs/ has the build logs (hardware / PNS / acoustics reports per file).

Trajectories and sequence diagrams

For each EPI: the (ky, kz) trajectory of frame 1 (physical k-space, m^-1; the start of the echo train is marked and the line connects consecutive echoes), and the first TR (RF, ADC and the three gradient axes, time in ms). The dashed line in the TR diagrams is the repo's nominal TE marker (echo ETL/2 - 0.5), not the k = 0 echo; the time of the k = 0 sample is in the table above (33.3 ms, 30 ms, 30 ms).

1_laminar_1x_2.4mm_ETL90: fully sampled, one ky line per train (frame = 60 trains)

Each color/number in the trajectory is one train: a ky line at one kz (60 trains per volume).

ky-kz trajectory, laminar

one TR, laminar

2_radial_90x_2.4mm_ETL60: R = 90, one train per frame (60 samples)

ky-kz trajectory, radial R = 90

one TR, radial R = 90

3_radial_122.7x_2.4mm_ETL44: R = 122.7, one train per frame (44 samples)

ky-kz trajectory, radial R = 122.7

one TR, radial R = 122.7

4_meGRE_dwell4us: first image excitation (22 echoes; dashed lines mark every echo center)

one TR, meGRE

Item 1 -- fully sampled laminar (1_...)

All-ones (ky, kz) mask, ETL 90 = Ny, 60 shots: each train is one full ky line (one kz, ky = 0..89 in unit steps, the same direction every shot), kz ascending across shots; no kz blips, only ky blips of one step. ArbEPI-python's own mask2epi_laminar does not give this (it fills each shot with consecutive samples of a kz-outer / ky-center-out raster), so generate.py swaps in mask2epi_ky_lines through sequences.ArbEPI._MASK2EPI; the repo is untouched. ky ascends in every train, so ky = 0 is on echo 45 of every line; the k = 0 sample (kz = 0, shot 30, echo 45) is at 33.301 ms after the RF center (measured from the played k-space trajectory; prescribed 33.3 ms, 2 µs is the 4 µs block raster).

Why TE 33.3 ms and not 30 ms. At ETL 90, with k = 0 on echo 45, the time before the first echo is 3.4 ms (half the water excitation, prephasers, dead times). At the 6 µs dwell the repo would pick, the earliest k = 0 echo is at about 33.3 ms. At the 4 µs dwell it would be at about 27 ms, so 33.3 ms (your choice) is reached by delaying the train by 6.3 ms. Volume TR is 60 x the minimum per-shot TR (62.83 ms), 3.77 s, so a 5-min scan has 80 volumes (301.6 s). Flip is the Ernst angle for the per-shot TR (17.7°).

4 µs dwell and the coil's acoustic band. ArbEPI's find_min_feasible_dwell rejects a dwell when the echo spacing is within 20 µs of one of the coil's forbidden acoustic-resonance bands (xrm: 410-510 µs on x/y, 360-440 µs on z). At 4 µs and the default slews arm 1's echo spacing is 524 µs, 14 µs above the 510 µs band edge, so the repo would have chosen 6 µs (660 µs); forcing 4 µs (FORCE_DWELL in generate.py, so that all EPIs share one receiver bandwidth) passed the repo's acoustics check with 0.2995 of the 0.3 limit, no margin. The readout ramp-up slew was therefore lowered to 120 T/m/s (blip 200, fall 190), which moves the echo spacing to 552 µs, outside the repo's band margin (>= 530 µs): acoustics 0.090 / 0.089 of 0.3, peak PNS 69.6 %, at +2 % per-shot TR (62.83 vs 61.60 ms). Chosen from a 12-point slew grid (pns_search-style, 1 frame, spoilers at maximum): lower ramp slews gave lower acoustics, e.g. 100 / 100 / 100: 640 µs, 0.027, +8 % TR.

Item 2 -- radial with the provided masks, one echo train per frame (2_..., 3_...)

mask_60.mat (60 samples) and mask_44.mat (44 samples) are byte-identical to 20260922xiaokai/seqs/mask_*.mat. Each frame is one echo train that traverses all samples of its mask: ETL 60 / 44, 1 shot, R 90.0 / 122.7, epi_trajectory = 'radial'. The frame TR is the minimum, 54.74 / 48.84 ms, so there are 5481 / 6143 frames in 5 min (a 2-shot split into two 30 / 22-echo frames needed 90 / 84 ms per pair). Flip is the Ernst angle (16.5° / 15.6°).

Same echo order every frame. The repo's current main chains the golden angle across frames, which makes the order differ every frame. Here the repo's own mask2epi_radial is run once in its per-frame mode (tau0 = None, the behavior of the 9/22 scans) with Nshots = 1, and that schedule is played on every frame (generate.py: use_arm replaces arb._compute_schedules; no repo edit). verify.py checks that scan_info.mat's schedules, including echo times, are identical in all frames.

k = 0 in every frame. The provided masks contain no k-space center, (ky, kz) index (45, 30) (the index ArbEPI's prephasers/blips treat as k = 0). mask2epi_radial puts the sample nearest the center on echo (ETL - 1) // 2 (29 for ETL 60, 21 for ETL 44); that sample is moved onto DC (your choice of 2026-11-11): mask_60's (46, 30) and mask_44's (45, 29). Sample count, ETL and R are unchanged. Each frame plays k = 0 exactly once, so there is no duplicate (ky, kz) within a frame. The collaborator should know each pattern changed by one sample. The edited sets are in scan_info.mat (schedules).

TE. ArbEPI-python's nominal TE sits at echo ETL/2 - 0.5, half an echo spacing off for even ETL, so TE is prescribed so that the DC sample's own kx = 0 crossing is at 30 ms (generate.py: design). TE is counted from the RF center that pulseq reports for the water-excitation pulse, the midpoint of the whole 1-3-3-1 train (2.112 ms into the 4.222 ms pulse; the amplitude centroid is identical). Checked on a built copy of the first frame (same timing as the full sequence) from the played k-space trajectory: the DC echo's kx = 0 crossing is 30.002 ms / 30.000 ms after the RF center, (ky, kz) = (0, 0) there to numerical precision. The residual is the 4 µs block raster. The saved per-echo times (scan_info.mat schedules(:,:,:,3)) give the same for every frame.

Dead time. With k = 0 on the middle echo of the train, TE 30 ms still needs a wait before the train: 7.8 ms (arm 2) and 13.7 ms (arm 3). Arm 3's train is shorter, so it waits more. Moving k = 0 later in the train would remove it but would break the spoke (a long jump in the echo order); not done.

Timing and PNS were measured on this schedule. Blip step maxima: ky 25 / kz 12 (arm 2), ky 19 / kz 11 (arm 3). At the repo's default slews (blip / readout fall / readout rise = 200 / 190 / 155 T/m/s) both arms are over or near the 80 % limit; slews were retuned with pns_search.py (coarse grid, 6 frames with every spoiler at its maximum, PNS <= 79 %): arm 2 160 / 160 / 140, arm 3 180 / 140 / 140, among the fastest per-shot TRs on that grid, picked for margin among near-ties. The real builds measure 76.5 % and 76.7 %. Arm 1 is at the defaults. The dwell is 4 µs for both radial arms (the repo's own choice).

Item 3 -- multi-echo 3D GRE (4_...)

sequences/meGRE.py (monopolar, one excitation reads 22 flyback-monopolar echoes per (ky, kz)), with the EPIs' FOV 216 x 216 x 144 mm, 2.4 mm isotropic, and their 4 µs ADC dwell (250 kHz receiver bandwidth), the same for all three EPIs, so there is one meGRE. Echoes sit at multiples of the fat off-resonance period (1/f_fat = 2.236 ms at 3 T, TE_n = n / f_fat, fat in phase with water): the first at 1 x 2.236 ms (the earliest multiple the prephasers fit in), up to the 22nd at 49.19 ms (a 23rd would be 51.4 ms > 50 ms). The 1-excitation-per-(ky, kz) scan (5670 shots including 2 dummy z-loops and the receive-gain block) takes 291 s. No deGRE was built (the repo still labels meGRE a prototype; a dual-echo deGRE can be added as in 20260922xiaokai/seqs/5_deGRE if a fallback is wanted). Each EPI folder's scan_info.mat carries the matching meGRE *_degre keys (Nx/Ny/Nz, fov, n_echoes, 22 echo times, TR, flip, dwell), so preprocessing reads them as a deGRE (main.py --gre megre equivalent). one_tr.png shows the first image excitation (iZ = 1, iY = 0) after the dummy z-loops and the receive-gain block, with every echo center marked. The readout is 27.2 mT/m (PNS 71.6 %, acoustics 0.155 of the 0.3 limit). meGRE.seq reaches 49.93 mT/m (limit 50), which is a gradient other than the readout.

Hardware (logs/)

File max grad (limit 50 mT/m) max slew (limit 200 T/m/s) max B1 (limit 0.25 G)
1_.../ArbEPI 27.18 188.8 0.0400
2_.../ArbEPI 27.65 158.0 0.0373
3_.../ArbEPI 27.39 177.1 0.0353
4_meGRE_dwell4us 49.93 174.3 0.1468

Running at the scanner

This is a suggested order, not something the collaborator specified. It follows the 9/22 session (20260922xiaokai/series_description.txt: GRE, then per EPI configuration a calibration scan, a noise scan and the EPI runs) and the repo's pipeline (noise -> whitening, EPIcal -> readout delay and odd/even phase, meGRE -> coil compression, sensitivity maps and B0).

  1. Transfer (done, 2026-11-12): the .pge files are on the inside scanner; type the entry number from the table below into the User CVs of a pge2 prescription (see ge/README.md in ArbEPI-python, "Running the sequence on the scanner console").
  2. Localizer / anatomical as usual (not part of these sequences).
  3. 4_meGRE_dwell4us/meGRE, once. It is shared by all three EPIs (same FOV, resolution and 4 µs dwell) and takes 291 s.
  4. For each EPI folder, in the order 1, 2, 3: EPIcal, then noise, then ArbEPI (5 min, one run: 80 / 5481 / 6143 frames). Use the same folder's three files together; each folder's scan_info.mat goes with its ArbEPI data in preprocessing.
  5. Prescription. Prescribe the same FOV / slab position for every scan (216 x 216 x 144 mm; the 9/22 session had a calibration scan that had to be discarded because the FOV was not prescribed).
  6. Data. Raw ScanArchives go in ../scanarchives/; note each Series number in the table below.

Series log of the 2026-11-12 ball-phantom session (Series numbers verified from each ScanArchive's readout count and length; entry numbers are the pge2 User CV values):

Order Scan File Entry # Series
-- localizer product -- 1
-- prescan (1 coil) -- -- 29000
-- MPRAGE product -- 2
1 meGRE (shared) 4_meGRE_dwell4us/meGRE 275 3
2 calibration 1_laminar_1x_2.4mm_ETL90/EPIcal 267 4
3 noise 1_laminar_1x_2.4mm_ETL90/noise 268 5
4 EPI 1_laminar_1x_2.4mm_ETL90/ArbEPI 266 6
5 calibration 2_radial_90x_2.4mm_ETL60/EPIcal 270 7
6 noise 2_radial_90x_2.4mm_ETL60/noise 271 8
7 EPI 2_radial_90x_2.4mm_ETL60/ArbEPI 269 9
8 calibration 3_radial_122.7x_2.4mm_ETL44/EPIcal 273 10
9 noise 3_radial_122.7x_2.4mm_ETL44/noise 274 11
10 EPI 3_radial_122.7x_2.4mm_ETL44/ArbEPI 272 12

The raw data (a ball phantom, not a human subject) is in ../scanarchives/Series<N>/; every noise Series holds a small ~2 MB prescan archive next to the real noise archive. Reconstruction output goes in ../recon/.

Reproduce

cd /home/rexfung/github/ArbEPI-python        # main, 28a28fa
.venv/bin/python /StorageRAID/rexfung/20261112comet_ball/seqs/pns_search.py radial_mask60  # slews (LIMIT=79 env); also radial_mask44
.venv/bin/python /StorageRAID/rexfung/20261112comet_ball/seqs/generate.py build           # folders 1-3 (all arms or named)
.venv/bin/python /StorageRAID/rexfung/20261112comet_ball/seqs/generate.py megre           # folder 4 (with one_tr.png)
.venv/bin/python -I /StorageRAID/rexfung/20261112comet_ball/seqs/verify.py                # checks scan_info.mat

generate.py measure [arm] (arms laminar_1x, radial_mask60, radial_mask44) prints each arm's timing without building anything.

Not done / flagged

  • Arm 1 is on a forced 4 µs dwell (the repo's own choice would be 6 µs); see Item 1 for the acoustics margin.

  • Scanner transfer (2026-11-12, ge/coppe.py --target inside, run id 20261112comet, run from a staging folder of uniquely named copies of the .pge files, since coppe takes one folder and every folder here has ArbEPI.pge, EPIcal.pge, noise.pge). Files are in /srv/nfs/psd/usr/psd/rexfung/20261112comet/ on the scanner:

    Folder ArbEPI EPIcal noise meGRE
    1_laminar_1x_2.4mm_ETL90 266 267 268
    2_radial_90x_2.4mm_ETL60 269 270 271
    3_radial_122.7x_2.4mm_ETL44 272 273 274
    4_meGRE_dwell4us 275

    Re-running coppe with the same run id would prompt before overwriting these entries.

  • recon/ (one level up) is empty. No human subject was scanned this session, only a ball phantom.

  • Items you may want to revisit: frame TR is the fastest per arm, no warm-up frames are discarded (discard_duration = 0), item 1's shots run kz ascending and echoes ky ascending.

About

Ultra accelerated single shot 3D EPI sequences, plus an accompanying multi-echo GRE sequence for Xiaokai's COMET project.

Resources

Stars

0 stars

Watchers

0 watching

Forks

Releases

Packages

Contributors

Languages