Postdoc position: Artificial intelligence and fast deep learning for proton Monte Carlo simulations
John Lee
Postdoc position:
Artificial intelligence and fast deep learning for proton Monte Carlo simulations
Job Information
· Organisation/Company: UCLouvain (Université catholique de Louvain)
· Department: Institut de recherche expérimentale et clinique
· Research Field: Computer science » Modelling tools, Physics » Mathematical physics
· Researcher Profile: Recognised Researcher
· Country: Belgium
· Application Deadline: 15 Feb 2024 - 23:59 (Europe/Brussels)
· Type of Contract: Temporary
· Job Status: Full-time
· Hours Per Week: 38
· Offer Starting Date: 1 Mar 2024
Offer Description
Monte Carlo (MC) method is a statistical methodology to estimate quantities by sampling and simulating the underlying events they are associated with. The academic example to illustrate MC simulations is to play darts in a circle to estimate p. The MC methodology is used in many fields, where simulating individual events is more accurate than a macroscopic model. In radiotherapy, MC methods are used to simulate interactions of radiation with matter to estimate the dose delivered to the patient. Different MC codes dedicated to PT exist, for examples MCsquare, FRED or GATE.
MCsquare is an open-source code used by IBA Dosimetry in myQA iON to perform an independent dose calculation for proton therapy. MCsquare is composed of transport algorithms and physical models to simulate both electromagnetic and nuclear interactions of ions with matter. Although nuclear interactions are less frequent than electromagnetic interactions, it is still mandatory to properly simulate them as they are responsible for the production of secondary particles, which affect the shape of the Bragg curve and the total dose (especially for the skin) by several percent.
Due to their lower probability, nuclear interactions are seldom sampled and thus accountable for most of the statistical noise in the resulting dose distribution. To get an acceptable statistical uncertainty with current MC dose engines like MCsquare, many events (i.e., protons) must therefore be simulated, which consumes time. In this paper, the authors needed to simulate 108 protons to reach statistical uncertainties of 1% and 1.5% for prostate and head and neck cases, respectively. The dose in water was computed in about 5 minutes.
Such simulation times are not suited to adaptive therapy or applications that require many treatment scenarios to be considered. A deliverable of the PSQUAD project aims at reducing the simulation time, while keeping an acceptable level of noise. For this purpose, several AI-based methods will be investigated. First a denoising neural network will be studied to smooth the noisy dose and thus to allow the user to drastically reduce the number of particles to be simulated. The second approach will consist in breaking down the MC engine into its electromagnetic and nuclear parts and to replace the latter with a Deep Learning model of nuclear interaction effects. Avoiding MC simulations for the calculation of nuclear interactions could drastically reduce the computation time of proton dose calculation. It will also allow us to develop a new dose engine for carbon therapy more easily by focusing the effort on the electromagnetic interactions and relying on the AI model for the simulation of the more complex heavy ion nuclear interactions, after training on reference models such as those available in Geant4. Several training strategies will be investigated, starting from regular training with manual collection of a relevant data set, and then going to automated strategies, closer to active learning or reinforcement learning, where the Artificial Intelligence (AI) model can increase its learning set by selecting representative/informative images in a large bank and querying the reference MC code (Geant4) to get the necessary dose (nuclear component).
MIRO (Molecular Imaging, radiotherapy, and Oncology) is a research center within IREC (Institute for Clinical and Experimental Research) of Université catholique de Louvain (UClouvain, Belgium). As a particle therapy research center, it has been a trusted partner for Ion Beam Applications (IBA s.a.), world leader in building and manufacturing proton therapy centers, for many years. Its expertise is recognized in the fields of radiotherapy, protontherapy and medical physics, leading to hundreds of publications. Relevant research areas of MIRO are:
- Clinical research protocols
- Reference dosimetry
- Advanced applications of Monte Carlo simulation in conventional radiotherapy and proton therapy
- Integration of functional imaging and tumour motion in adaptive intensity modulated therapying non-small cell lung cancer
- Artificial intelligence in radiation oncology applications
- Robustness optimization and evaluation of proton therapy treatment plans
The open position is for an experienced researcher in one of the targeted fields (applied CS or physics, biomedical engineering, etc.). A postdoc researcher will be favoured, although the 2+2 funding could also accomodate a less experienced researcher (PhD student). MIRO has strong contacts with ICTEAM (engineering institute of UCLouvain).
UCLouvain is one of the top-level French-speaking universities of Belgium, with this position located on the Brussels campus.
Requirements
· Research Field: Computer science » Modelling tools
· Education Level: PhD or equivalent
· Research Field: Physics » Computational physics
· Education Level: PhD or equivalent
· Skills/Qualifications: Good computer programming skills, good knowledge of AI/DL/DNN/CNN, good knowledge of applied particle physics (medicine) and/or computational physics (Monte Carlo simulations).
· Languages: ENGLISH
· Research Field: Computer science » Modelling toolsPhysics » Computational physics
· Years of Research Experience: 1 - 4
Work Location(s)
Company/Institute: UCLouvain/SSS/IREC/MIRO, Brussels 1200, Belgium
Contact
Submit your application through Euraxess.