1. A Brachytherapy Treatment Planning Software Based on Monte Carlo Simulations and Artificial Neural Network Algorithm
Amir Moghadam

2. Prostate BrachyTherapy

3. Treatment Planning System(TPS)
A Treatment Planing System essentially consists of three parts:
a system for dose prescription.
a set of rules to distribute the sources inside a defined volume to achieve a clinically acceptable dose distribution.
a method to calculate patient dose.

4. Dose Calculation Methods
Analytical models (not used at low energies)
Deterministic solution to the transport equation (almost fast, also accurate)
Monte Carlo (most accurate but too slow)
TG-43 Formalism (fast but not accurate)
assumes human body as homogeneous water
As of 2009, the current approach for brachytherapy dose calculation is based on the AAPM TG-43 dosimetry formalism, which relies on superposition of single-source dose distributions obtained in a liquid water phantom with a fixed volume for radiation scattering.
pencil beam, Cone Colaps, and superposition convolution. Advantages of CC and superposition-convolution implementations are their availability as TPS dose calculation algorithms. Furthermore, they can accurately account for material inhomogeneities in external beam radiation therapy through clinically validated algorithms.
Problems of Analytical Methods
Photon energies are lower in brachytherapy than for external
beam. As such, the radiation path length is shorter and
the contribution of scattered radiation to the dose, at relevant
distances, can be of the same order of magnitude as the primary
dose. Furthermore as photon energy decreases, the
amount of energy released per interaction decreases and multiscatter
quickly becomes the dominant contribution of the
total scatter component.
Deterministic solution to the transport equation
Another calculation approach consists of directly solving the linear Boltzmann transport equation LBTE through deterministic means.
Discretization is also performed in space finite difference or finite element and in energy with appropriate multigroup cross sections.
It is generally understood that the accuracy of deterministic approaches is directly related to discretization, with fine steps leading to accurate solutions at the price of a larger system of equations to be solved
Zhou and Inanc102 introduced another approach to solve the integral representation of the LBTE and studied ISA in 125I seed implants. The algorithms were extended to account for fluorescence x rays, and parallelization was introduced. Single seeds and more complex configurations such as the Quality Assurance Review Center geometry for clinical protocol accreditation were studied. The calclation took 2.5 h on a 64 processor computer cluster
Monte Carlo
Direct MC simulation is based on a random sampling of particle histories to estimate the quantity of interest, absorbed dose in the patient.
2007 One such study has argued strongly in favor of MC dose calculation as a clinical standard for postimplant dose evaluation.

5. What are we going to do?
We want to create a method which is as fast as TG-43 Formalism but has a similar accuracy to Monte Carlo method.
To do this, first we have to create a 3D model of the patient’s body in our Monte Carlo code MCNP, based on his CT Scan image.

6. How to define the material and densities
CT Number to
to 90
91 to
Material
Air
Soft tissue
Bone Cortical
Density range (g/cm^3 ) to to to

7. Dicom to MCNP input conversion

8. Two methods of MCNP calculation
Total phantom Modeling
Supper position Method

9. The reason for Dose calculation by ANN
Dose calculation using MCNP takes at least 4 hours to reach an acceptable error on a Cori-7 computer.
To improve the speed of Monte Carlo dose calculation method we trained a set of Artificial Neural Networks to perform the dose calculation in a shorter time but with the accuracy similar to MCNP calculations.

10. How ANN works
Inputs : attenuation coefficients of the voxels of the cube
Output: 3D Dose distribution inside the cube because of the seed at the center of the cube

11. The ANN Training Flow cchart

12. Last step (using data compression)
Outer region 0.6cm 10*10 mesh
Inner region
0.3 cm 8*8 mesh
Model details are reduced as we get far from the seed

13. Time and number of sample cubes used for training
We modeled model cubes inside MCNP and we used of these model cubes for training the ANN and the remaining were saved for validation of ANN for unknown cases.
MCNP calculations took 40 minutes for each of the model cubes.(total MCNP simulation for cubes will be103 days on 12 Nodes!!?)
Training of each of ANNs took about 20 to 60 minutes.

14. Results of inner region for single seeds

15. Results of inner region for single seeds

16. Results of inner region for single seeds

17. Results of outer region for single seeds

18. Results of outer region for single seeds

19. Results of outer region for single seeds

20. Validation of ANNs

21. Realistic Treatment Plan

22. Non-realistic treatment plan

23. Speed assessment of the New Method
Dose Calculation Method
Accuracy
Time needed for 1 seed
Time needed for 467 seed positions
MCNP
Highest
40 minutes for 1 seed
12 days on a cori-7 computer
Varian’s Acuros
(Linear Boltzman Transport Equation Solver)
High (dependant on the mesh size)
3-8 minutes for one change in position
23 hours
Artificial Neural Network
Similar to MCNP
8 minutes for 1 seed
17 minutes on a rusty computer

24. Future work
This study can be extended to the cases of High Dose Rate and Low Dose Rate BrachyTherapy and any other field in which effects of heterogeneity are much more of interest and the speed of calculation is critical.
Other methods of compression can be tested for improving the ANNs accuracy.
Other Architectures of ANN can be used for decreasing the relative errors of ANNs for Farther voxels.

25. Thanks for your attention