1. Geant based inverse planning for radiotherapy Low-hanging fruit
Balazs Ujvari
Hungary

2. Geant based inverse planning
Motivation
Computing for medical applications is a strategic domain with a wealth of possible projects that will be assessed by CERN’s medical applications decision-making structure. (Strategy p13)
Geant is mainly used for studies in the radiotherapy.
Geant based inverse planning

3. Why are these differences?
They use sophisticated methods to calculate dose at given position
BUT works well in homogeneous volumes
The validation is done in homogeneous phantom with air-filled ionization chamber
Geant based inverse planning

4. What can cause these differences?
Since the surface of the treatment is not well known
They add some extra to burn out the tumour
Depending on the device, geometry and the neighbouring organs it’s about 0.5 – 1mm
Healthy organs can be damaged
Geant based inverse planning

5. Our work at University of Debrecen
Dose deposit balls (shots) 4/8/14/18 mm
Geant based inverse planning

6. Geant based inverse planning
What we can do
Start always from scratch
Build the Geant4 volume from CT image
(~70M voxel = G4Box)
Analyse the geometry of tumour
Make a list of reasonable 4/8/14/18 shots
Run the Geant4 on local supercomputer (can use max 500 cores)
10k runs (sometimes 100k) few days – one week
The results are the ROI (dose in 5-10M voxels)
Collect all the shots and try to put inside the volume (100GB raw data)
Find the best shots to fill the volume (and minimise the side effect)
Naive AI methods, search trees
As result there is a list with the central and the diameter of the shots, it can be converted to rotations
Geant based inverse planning

7. Geant based inverse planning
What they need
A button in the
menu
Run MC
Inverse Planning
NOW
Nice study
Useful
Something great
Geant based inverse planning

8. Geant based inverse planning
Plan to be useful
Show that we can do it faster (1-4 hours)
Run Geant4 for thousands of CT images to enhance the preselection
Learning the similarities perhaps 1k or less runs (shots) would be enough
Short runs for candidates to exclude the obvious wrong results
Cloud/grid computing with the best candidates
Better methods to find the ones, that can fill the tumour with minimal side effect, perhaps without move the results in the local computer
Help the traditional methods to handle inhomogeneous body
CT images, geometry of the tumours
Run at first on supercomputer
Think about cloud/grid solutions
Check the recent AI solutions for the filling problem
Nice study
Useful
Something great
Geant based inverse planning

9. Plan to do something great (in 5 years)
Discuss about the tendencies in cancel treatment
What will be the top100 geometry in 5-10 years
Shot sizes should be parameters
4 / 8 / 14 / 18 mm what kind of optimisation is it?
3.58 / 6.29 / / ??
Run thousands of CT with thousands of shot combinations
Talk with the R&D at the companies (hardware and software)
Central DB from the treatment to learn from the past
Help to develop the next devices, find best treatment
Estimate the cost to be realistic (cloud 10 x 16cores 2 hours 10CHF)
Nice study
Useful
Something great
Geant based inverse planning

10. Plan to do something great (in 5 years)
Find the fastest method to fill the volume
Prepare for the more detailed medical images (0.5mm → 0.2mm)
100M voxels → 1G voxels, or for half/full body 10-20G voxels
Moore-law won’t help us, the CPU need for the calculation with detailed images could grow faster
CT images, geometry of the tumours
Work together with Geant V developer
Dedicated cloud/grid development
Company R&D, know the tendencies
Much better search methods
User friendly, integrated solutions
Nice study
Useful
Something great
Geant based inverse planning

11. Geant based inverse planning
Thanks!
Geant based inverse planning