PhD project: Machine Learning and FPGA optimisation for proton beam therapy

Supervisor: Dr Simon Jolly

Proton therapy is a more precise form of radiotherapy that provides significant benefits over conventional X-ray radiotherapy, particularly for children. Two new detectors currently under development within the HEP proton therapy group are focussed on improving Quality Assurance (QA) in proton therapy, but have applications in patient positioning and imaging. All of these applications will assist in making proton therapy safer and more accurate for patients as well as allowing more patients to be treated through saved QA time. The QuARC — currently funded by a 3-year STFC IPS grant — measures the proton range with a series of plastic scintillator sheets. The scintillation light output is measured by a series of photodiodes, which are read out by custom ADC boards controlled by an FPGA. The QuADProBE — recently funded by a 3-year STFC CLASP grant — adds beam size and position through orthogonal arrays of scintillating fibres, again read out by photodiode arrays and controlled by FPGA. The intention is to include dose measurements to the final clinical system, thereby providing a single detector capable of making all 4 clinical beam QA measurements at once: this would be the first such clinical system.

This project will focus on the reconstruction of the proton beam range, position and size by FPGA. The FPGA DAQ code has already been written, but the fitting must be carried out in software, placing more strenuous requirements on the back end PC controller. By integrating the fitting onto the FPGA, a smaller, lighter, more cost effective detector can be developed. The student will need to learn how the existing FPGA code works for the range DAQ, then extend this to include the beam size and position and incorporate the fitting routines. Machine Learning techniques will also be applied to the fitting procedure to improve the accuracy of the 3D reconstruction: this is particularly critical for the Birks constant fit that defines the amount of scintillator light quenching and also enable the reconstruction of Spread Out Bragg Peaks (SOBPs), which is not currently possible using existing reconstruction techniques with scintillators. This will necessitate the adaption of methods from Machine Learning, including non-parametric probability models and deep-learning techniques. Extensive detector simulations will be developed to enable training of the ML algorithms which will then be applied to data taken during previous experimental runs as well as online live reconstruction during subsequent clinical beam experiments.

The project will be jointly supervised by Dr Simon Jolly and Prof. Adam Gibson. Simon Jolly is a recognised expert in the fields of proton therapy accelerators, detectors and diagnostics with 20 years’ experience as an accelerator physicist and the founder of the current proton therapy research programme within the UCL HEP group. Adam Gibson’s main area of expertise is optical imaging in medicine has collaborations with hospitals in London, Cambridge and Birmingham. He was also one of the founders of the UCL Medical Physics proton therapy research group. It is also anticipated that a member of the UCL DIS CDT with specialist knowledge in Machine Learning techniques will join the supervisory team.

The first year of the PhD will be primarily be dedicated to training in the requisite areas: this will include the necessary underlying particle and medical physics knowledge — as well as Machine Learning techniques as part of the CDT — but will focus on FPGA programming and implementation of fitting algorithms on FPGAs. During the second year, the student will shift focus to building advanced detector simulations of both detectors that will provide the input for training the ML algorithms for fast range and position reconstruction. It is also anticipated that the student will contribute to the DAQ development and data analysis of any experimental tests. By the third year the majority of the student’s time will be dedicated to implementation of the improved reconstruction and fitting algorithms on the detector controller FPGAs as well as development of the web-based front end control GUI: this will continue into the fourth year as the FPGA implementation broadens in scope and becomes more sophisticated — combining all parts of the measurement into a single unified display — in addition to the increased availability of experimental data.

More details on the project can be found on the UCL HEP Proton Therapy page.

For questions about the project, please contact Dr. Simon Jolly.