Spiral Simulation Running out of GPU Memory

[nathanjbeaumont]

Hello,
I have a spiral sequence with 48 interleaves I'm trying to simulate with the brain_phantom2D. To get an image that doesn't have the screen door looking artifact I need to set us=2 in brain_phantom2D(). However if I do this even with only 48 interleaves in the sequence my GPU runs out of memory. I have an RTX 2070 16gigs of GPU memory, (8 dedicated, 8 shared). Run tests.jl in the zip below to see:

tests.zip

I'm able to run it on CPU but the recon looks very bad for a fully sampled k-space:

[cncastillo]

Hi!

Regarding the GPU running out of memory, you can reduce the size of the simulation blocks by increasing sim_params["Nblocks"]. So for example:

sim_params = KomaMRICore.default_sim_params()
sim_params["Nblocks"] = 2_000

Regarding the reconstruction, If I reconstruct the first 16 readouts, I get:

obj = brain_phantom2D(; us=2) 

I chose that number because the signal appears to have some oscilations:

I believe this is happening because of the flip angles being too high for the acquisition type. If you are trying to create a balanced sequence, there are a few things missing:

• Add a refocusing gradient in z-direction
• Alternate flip angles (+alpha, -alpha, +alpha, ....).

I patched the sequence by doing this (only the second point):

seq = read_seq("temp.seq")
idx = findall(is_RF_on.(seq))[2:end] # Removes IR pulse from phase alternation
for (i, idx) in enumerate(idx)
    factor = ((-1)^(i-1) + 0im)
    seq.RF[idx] = factor * seq.RF[idx]
    seq.ADC[idx+3].ϕ = angle(factor) # The phase factor in the ADC is important
end 

and I get something that looks much better

As a side note I would avoid to use a slice-selective RF pulse for the initial inversion, it is much better to use an adiabatic hyperbolic secant inversion. Let me know if you need help implementing it, but there are some examples in the examples folder.

[nathanjbeaumont]

Thanks for the response.

Add a refocusing gradient in z-direction

Do you mean a rewinder gradient in z after the RF pulse? There is already a rewinder after every RF pulse:

Alternate flip angles (+alpha, -alpha, +alpha, ....).

The paper I'm trying to reproduce uses what they are calling "FISP". and they say:

The long-used Fast Imaging with Steady State Precession (FISP) sequence acquires coherent steady state signals with a constant unbalanced gradient moment in each repetition time

So I was assuming they used the a phase of 0 for every RF pulse. However this article claims FISP uses alternating pulse phases. This article also says:

rapidly applied RF-pulses result in saturation of the longitudinal magnetization. If the RF-pulses were continuously applied in the same (+α) direction, this steady-state magnetization is lower

[cncastillo]

Oh sorry, I thought you wanted to recreate a trueFISP, aka balanced Steady State Free Precession (bSSFP). https://pmc.ncbi.nlm.nih.gov/articles/PMC3602925/

The additional gradient in z can be seen here:

In that case, with MRF-FISP you generally use spoiled gradient echo (https://www.sciencedirect.com/science/article/pii/S1936878X18308337):

The cMRF pulse sequences used to date have a fast imaging with free precession (FISP) readout with no radiofrequency phase cycling, instead of the originally proposed bSSFP readout. Unlike the bSSFP readout, FISP uses an unbalanced spoiler gradient in the slice select direction, thus making this sequence relatively insensitive to off-resonance effects.

For the simulation, to have a realistic gradient spoiling, you need to add spins in the z dimension or use brain_phantom3D (you can use the start_end keyword argument to adjust the thickness and us for the up-sampling).

[nathanjbeaumont]

The paper also says this:

The FISP sequence does not lead to the banding artifacts that are seen in bSSFP. Without any other mechanisms to destroy the coherence of the transverse magnetization (such as RF spoiling), the sequence is sensitive to both longitudinal and transverse relaxation times.

So it sounded to me like they didn't use spoilers. They're purposefully trying to keep transverse magnetization. Also you can see they rewind back to the center of k-space after the spiral, which would be pointless if they're going to spoil right after the rewind right?

[aTrotier]

Take a look at this youtube video : https://youtu.be/6wAuxKf1GhQ?si=UQ06l2gw9kKJe_Ib

The "spoiler" needs to be constant for each TR (and no RF spoiling) in order to obtain a FISP sequence

[nathanjbeaumont]

OK I believe I figured out the issue, the paper says:

Figure 1a shows one example of the MRF-FISP sequence with the unbalanced gradient along the slice-selection direction. In this study, we used a SINC waveform RF pulse with 2 ms duration. The time bandwidth product of the RF pulse was set to 8 to minimize the imperfect slice profile. The unbalanced gradient moment in the slice-selection direction achieved 8π dephasing across 5 mm slice thickness within each repetition time.

So I believe the rephasing gradient after slice selection had an area less than half the slice selection gradient so that the rephasing was incomplete (8π of dephasing remaining).

[aTrotier]

No they said that their is a net moment of 8pi at the end of the sequence along the slice direction. You must rephase everything along read and phase and just put an additional spoiler at the end of the TR along the slice dimension of 8pi area. If you put the de phasing gradient of 8pi after the slice selection you will obtain something like a PSIF sequence I think.

[cncastillo]

This! ⬆️

An $8\pi$ spoiler along the slice direction is what would make sense to me. The phase dispersion of this spoiler is dependent on the slice thickness $\Delta z$, so you need to calculate the area of the gradient $A_{\mathrm{grad}}$ such that

$$ \phi = \gamma A_{\mathrm{grad}} \Delta z = 8\pi. $$

[nathanjbeaumont]

@cncastillo I would like to use an adiabatic pulse for the inversion at the beginning as that's what the paper does, but I don't know how to design such a pulse. I see examples\3.tutorials\adiabatic_pulses\BIR4.mat. How was BIR4.mat generated? Is the script for that available?

[cncastillo]

I think I lost the script that generates it unfortunately, but at the top of the scripts I put the name of the paper used as reference. I believe that in the paper you are trying to replicate they use a hyperbolic secant (HS) pulse. These are some examples:

• BIR4: https://github.com/JuliaHealth/KomaMRI.jl/blob/master/examples/3.tutorials/adiabatic_pulses/BIR4InversionProfile.jl
• HS:

  KomaMRI.jl/examples/4.reproducible_notebooks/pluto-03-low-field-boost-optimization.jl

  Lines 188 to 215 in 6d4deb8

  	function IR(IR_delay; sample=false)
  	# Generating HS pulse
  	# Based on: https://onlinelibrary.wiley.com/doi/epdf/10.1002/jmri.26021
  	# Params
  	flip_angle = 900; # Peak amplitude (deg)
  	Trf = 10240e-6; # Pulse duration (ms)
  	β = 6.7e2; # frequency modulation param (rad/s)
  	μ = 5; # phase modulation parameter (dimensionless)
  	fmax = μ * β / (2π); # 2fmax = BW
  	# RF pulse
  	t = range(-Trf/2, Trf/2, 201);
  	B1 = sech.(β .* t);
  	Δf = fmax .* tanh.(β .* t);
  	# Spoiler length
  	spoiler_time = 6000e-6
  	spoiler_rise_fall = 600e-6
  	# Prepulse
  	seq = Sequence()
  	seq += RF(B1, Trf, Δf) # FM modulated pulse
  	seq = (flip_angle / get_flip_angles(seq)[1] + 0.0im) * seq # RF scaling
  	seq += Grad(8e-3, spoiler_time, spoiler_rise_fall) #Spoiler3
  	if sample
  	seq += ADC(11, IR_delay - spoiler_time - 2spoiler_rise_fall)
  	else
  	seq += Delay(IR_delay - spoiler_time - 2spoiler_rise_fall)
  	end
  	return seq
  	end

• HS: https://github.com/JuliaHealth/KomaMRI.jl/blob/master/examples/3.tutorials/adiabatic_pulses/HSInversionProfile.jl

To be honest some of this are almost ready to be a Literate example. @Stockless could you create one based on HSInversionProfile.jl?

[cncastillo]

I found some Mathematica code to generate it that might be useful:

You may want to adjust the $\beta$ and $\kappa$ parameters.

[aTrotier]

I am using the sigpy python package for that (already implemented in pypulseq)

[aTrotier]

Take a look at : https://sigpy.readthedocs.io/en/latest/_modules/sigpy/mri/rf/adiabatic.html#bir4