Fit Proton Bragg Peaks

Fit Bragg peaks with bortfeld function

Proton Bragg Peak Fit

Feature

1. Fit IDD curve with Bortfeld functions, IDD curve may contains multiple bragg peaks.
2. Reconstruct 2D/3D proton dose from measurement of MLSIC^[5], (multi-gauss model in XY plane, bortfeld in Z)
3. Proton Radiography/CT with feature 1 and 2

---

• Provide IDD measurement data and simulation data, data acquired from Zebra and MLSIC which can be used in Proton Radiographics
• Bortfeld function implemented in C++, provide IDD, mean gradient and jacobian outputs.
• Input Integral Depth Dose(IDD) is suggested to rescale to [0,10] or nomalize to [0,1], but it's ok to fit raw data in any scale
• Compiled with Visual Studio + Intel OneAPI, faster than mex in MATLAB. Highly recommanded to compile the src with VS and Intel OneAPI
• Support Apple Silicon (Matlab R2022a, Matlab R2022a beta don't have toolboxs)
• Image Processing Toolbox, Optimization Toolbox and Signal Processing Toolbox required

Working in progress, bragg-peak fitting are ready to use.

How to use:

Windows

• Pre-compiled functions are provided.
  • BortfeldFunction.mexw64 Windows with avx2(supported by most of modern x86_64 CPU).
• If Pre-compiled functions not work, run ./src/compile_PBPF.m to compile it

MacOS

• Pre-compiled functions are provided.
  • BortfeldFunction.mexmaca64 for Apple Silicon Mac (do not Support AVX)
  • BortfeldFunction.mexmacai64 for Intel Mac (AVX not used)
• If Pre-compiled functions not work, run ./src/compile_PBPF.m to compile it

Linux

• Pre-compiled functions are Not provided.
  • BortfeldFunction_avx512f.mexa64 Linux with avx512f (Only some x86_64 CPU support this) If you use this, change BortfeldFunction in bf_mex to BortfeldFunction_avx512f
  • BortfeldFunction.mexa64 Linux with avx2(supported by most of modern x86_64 CPU)
• If Pre-compiled functions not work, run ./src/compile_PBPF.m to compile it

Compilation

• All precompiled mex function can work with both single and double, but single numbers can not be used for finite difference optimization(double precision required). Use the gradient provided by mex function.
• Open the root folder run cd('./src') in matlab
• run compile_PBPF three compile flags are provided
  • use_openmp = 1; Use OpenMP
  • use_avx = 1; Use AVX2
  • is_avx512 = 1; Use AVX512 (not every CPU support this)

Here are some code snippets

• predict IDD with depth and bf parameters

z = (1:64)*0.3; % depth unit(cm), (n,1) size is preferred
% bortfeld function parameters, [range1,sigma1,epsilon1,phi1, range2,sigma2,epsilon2,phi2] two bragg-peaks used here, range unit is cm
bf_para = [15,0.3,1e-3,0.4, 12,0.4,1e-3,0.4];
% get idd curve, size of output will be (n,1), no matter z is (1,n) or (n,1)
output = bf_mex(z,bf_para,'idd');
plot(z,output);
grid on;
grid minor;
xlabel('Depth(cm)');
ylabel('Dose(a.u.)');% a.u. means arbitary unit
% for short
% output = bf_mex((1:64)*0.3,[15,0.3,1e-3,0.4, 12,0.4,1e-3,0.4],'idd')

• Fit measured IDD

% method 1
[x,idd_o] = precise_fit(z,idd_i,num_bp,strict);
% method 2 used for fast fit, don't calculate IDD in this function
x = fast_fit(z,idd_i,num_bp);
idd_o = bf_mex(z,x,'idd');

• Get gradient (Jacobian matrix) of Bortfeld function

z = (1:64)*0.3; % depth unit(cm), (n,1) size is preferred
% bortfeld function parameters, (m,1), range unit is cm
bf_para = [15,0.3,1e-3,0.4, 12,0.4,1e-3,0.4];
% output size will be (n,m)
output = bf_mex((1:64)'*0.3,[15,0.3,1e-3,0.4, 12,0.4,1e-3,0.4],'jacobian')

• Other functions
  • stored in ./utils/
  • If matlab crashed on Rosseta Matlab, run this function ifMacCrashed()
• Benchmark of mex functions ./utils/test_mex_func.m, further optimization required.

% BortfeldFunction time: single(0.249554s), double(0.248533s)
% BortfeldFunction Grad time: single(2.758517s), double(2.772066s)
% Gauss2D Grad GPU time: single(0.942952s), double(1.575574s)
% Gauss2D Grad CPU time: single(0.131198s), double(0.201214s)
% Gauss2D GPU time: single(0.196284s), double(0.450232s)
% Gauss2D CPU time: single(0.842755s), double(0.665228s)
% ProtonDose3D GPU time: single(0.493263s), double(0.537123s)
% ProtonDose3D CPU time: single(0.751850s), double(0.496489s)

s5cuGAUFbX1a

1D IDD data

Applications can be found in ./utils/demo.m and data stored in ./data/. Be careful, run the demo section by section, some parts are time comsuming.

Zebra data	MLSIC data

2D dose reconstruction

Strip ionization chamber array[4,5] were used for QA purpose, two layers of horizontal and vertical strip detectors were used to locate the proton beam position.Details

White Noise level(divide by max dose)	Max Absolute Error( <math-renderer class="js-inline-math" style="display: inline" data-static-url="https://github.githubassets.com/static" data-run-id="aedc40f729f35be5b8c316392f7acbaa">$\frac{\max|error|}{max[dose]}$</math-renderer>)
30%	0.0441
20%	0.0337
10%	0.0192

3D dose reconstruction

Only MLSIC can measure a 3D dose in real time[5], it contains 768 ADC channels (XY: 128x2, ZX: 8x32, ZY: 8x32) Details. The measurement function of MLSIC can be defined as a function, <math-renderer class="js-inline-math" style="display: inline" data-static-url="https://github.githubassets.com/static" data-run-id="aedc40f729f35be5b8c316392f7acbaa">$f: \R^{128\times128\times66} \rightarrow \R^{128\times2 + 8\times32 + 8\times32}$</math-renderer>

Symmetric dose

3D Dose	MLSIC measurement

Central Slice Dose	IDD

Asymetric dose

to be updated

Proton Radiography

Gammex Phantom	MLSIC data

Head Phantom	Simulation	MLSIC data

Proton CT

to be updated

Details of math and MLSIC structures can be found in here

Reference :

1. An analytical approximation of the Bragg curve for therapeutic proton beams
2. E. Cojocaru. Parabolic Cylinder Functions
3. On the parametrization of lateral dose profiles in proton radiation therapy
4. A 2D strip ionization chamber array with high spatiotemporal resolution for proton pencil beam scanning FLASH radiotherapy
5. A multi-layer strip ionization chamber (MLSIC) device for proton pencil beam scan quality assurance
6. Evaluation of the influence of double and triple Gaussian proton kernel models on accuracy of dose calculations for spot scanning technique

Remark:

M1 mac openmp issue: https://www.mathworks.com/matlabcentral/answers/1761950-m1-mac-compile-mex-file-with-openmp?s_tid=srchtitle

Matlab R2022a thread safty is not very stable, if the mex function crashed matlab, run the magic function 'test3_mex.mexmaci64' by calling test3_mex(int32(16));. Its output looks like

i = 1 
i = 2 
i = 7 
i = 8 
i = 5 
i = 6 
i = 3 
i = 4 
i = 9 
i = 10 
i = 13 
i = 11 
i = 14 
i = 15 
i = 12 
i = 16 

Matlab R2022a beta works well with openmp but it do not support official toolbox.