Multistage flagging system for online adaptive proton therapy

Lately the high clinical potential of proton therapy, coupled with the appearance of new technologies such as pencil beam scanning and the mounting of imagery tools in the very midst of the treatment room increased the interest around its development. This particular cancer treatment modality takes advantage of the singular way the protons deposit their energy throughout the matter: the Bragg peak. This very narrow dose profile peak allows very little dose deposition before the peak and close to no dose after. But as efficient as it is, this energy deposition is also very sensitive to density variations and if we want to know exactly where the peak will be located we have to developed methods to account for these variations. One of them consists in an online checking for these variations by imaging the patient anatomy in the tumor area right before his daily dose delivery to make sure it still corresponds to the images taken for the treatment planning. If there is a mismatch we can choose to adapt his treatment to ensure a good dose conformity and a safe dose delivery. This is what is called online adaptive proton therapy. That is in that context that our research question was raised. How can we determine if a patient needs a treatment plan adaptation or not? Of course we could compute the dose distribution every time the patient lay on the table for its treatment but the dose distribution is way to long to compute. In fact, the more we wait for approval, the more likely the patient is to move, leading then to new positioning errors. We therefore need a flagging system working in the order of the second. We decided this system would work on 3 different stages of increasing complexity. The first stage, and the fastest one, would be based on the difference evaluation between the reference scanner and the daily scanner using Hounsfield unit (HU) matrices. The second stage would consist in computing the difference between WEPL maps from both reference and daily scanner. And finally, the third and last stage, which is also the most time consuming one would be the computation of the dose distribution. For this purpose, using a pool of 13 patients, we tried to find metrics that could serve as adaptation indicators. We started with the dose distribution based indicator as we wanted to use it as a reference to compare our future metrics with. We found the homogeneity index to be performing very well and we decided to use it. We then came to the $\Delta HU$ based indicators and we found that peak signal to noise ratio and the mean squared error could be used as a first way to filter out patients not needing a treatment plan adaptation. For the WEPL based indicators, the percentile 90 performed very satisfyingly, allowing us to discriminate with large margins both patients to adapt and patients not needing adaptation. Our study showed us that a multistage flagging system for online adaptive proton therapy was feasible even though we could improve the robustness of our conclusions if we had more patients.