Machine learning model for radiation dose
My Topol fellowship problem / project:
A patient’s radiotherapy treatment is carefully designed to deliver optimal dose distributions which will maximise tumour control whilst minimising the probability of normal tissue toxicity. This is achieved using commercial treatment planning systems that utilise proprietary analytical dose calculation systems which are very fast. This is a necessary characteristic as treatment planning is a highly iterative process that can take days to achieve a clinically viable plan.
However, it is widely accepted that a simulation approach called Monte Carlo is the gold standard for accuracy. Monte Carlo is not used for planning as it is prohibitively computationally intensive taking many hours to days to calculate a single iteration and is, therefore, either entirely omitted or reserved purely for final dose checks after the planning process has been completed.
The issue with not being able to utilise Monte Carlo during the planning process is that the inaccuracies of the commercial treatment planning system must be acknowledged by the clinicians when reviewing a plan. In proton beam therapy there are two groups of patients that are of specific concern where quantitively accounting of these inaccuracies would be especially impactful. Firstly, patients who have high-density surgical implants experience increased proximal dose to the implant.
However, commercial treatment planning systems do not correctly model sharp changes in density and thus under-represent this phenomenon. Monte Carlo would allow for a more accurate depiction of these effects and provide confidence to the reviewing clinician of the dose delivered to neighboring sensitive organs (e.g., spinal cord).
Secondly, Monte Carlo can also provide other physical metrics which describe how the radiation is dosing the tissue. One such metric is the linear energy transfer which is known to alter the radiobiological effectiveness of administered dose. Both dose inaccuracies and lack of linear energy transfer calculation result in clinicians having to make a clinical judgement without tangible quantitative values. This can result in requiring treatment plans to be re- made or if there are safety concerns it may result in the decision to decrease the dose to avoid potential normal tissue toxicity. If Monte Carlo dose distribution and linear energy transfer can be modelled at the time of planning these critical decisions will be better informed.
This project aims to achieve this by leveraging newly developed machine learning models that can rapidly predict Monte Carlo dose and linear energy transfer on patient plans. The ability to access Monte Carlo predicted data at the planning stage will empower planning staff to deliver more quantitively defined clinical treatments. This project will facilitate the clinical implementation of the models, which includes the independent validation of the models, interfacing with the clinical treatment planning system, creation of procedures, risk assessments, workflow design and departmental training. It will involve working closely with clinicians, physicists, and dosimetrists to ensure all stakeholders have been involved in achieving a clinical solution. If successful we can then share what we have learnt whilst implementing the clinical solution with other proton centres at a national and international level.
Clinical Scientist specialised in radiotherapy physics. Working at the Proton Beam Therapy Centre based at The Christie NHS FT, Manchester. A dual role within the clinical and research proton therapy team.
Interested in how we translate from research through to the clinic.