Ponente
Descripción
Proton therapy is a radiation treatment that targets tumoral cancers more precisely than conventional radiotherapy. This is because most of the dose is deposited near the end of the proton range at the Bragg Peak. However, different factors such as gas regions in the body, patient movement, or short-term physiological changes can produce deviations in the dose deposition. In turn, this may lead to tumor undertreatment or damage to crucial organs. Therefore, it is crucial that these deviations are detected in a proton therapy treatment.
Among other methods, it is possible to use positron emission tomography (PET) to estimate the dose from the generated positron emission activity that the proton beams induce. To this end, a recent work from our group proposed using a precomputed patient-specific and treatment-specific Dose-Activity Dictionary (DAD). With this DAD, the positron activity measured using a PET scanner immediately after the treatment could be translated into a dose distribution. This method searches for the linear combination of cases that best reproduces the observed activities. However, the performance depends significantly on how the considered case is similar to the simulated ones.
In this work, we propose a model based on a 3D U-Net architecture that accurately predicts the deposited dose from the activity induced by new, non-simulated proton beams. The U-Net can do this by capturing the high- and low-level features relating activity to dose at different tissues and proton beam spot depths. This model consists of three encoder and decoder layers and is trained on 948 simulated activity-dose pairs from the DAD. The training is completed in 20 minutes using an NVIDIA V100 GPU, which is sufficiently fast considering that it would be trained on data from a CT scan obtained a day before the treatment.
Preliminary results show a similar relative error distribution with respect to the previous work, as well as range deviations under a millimeter. Moreover, no signs of overfitting have been observed. Other recent methods, such as diffusion models and vision transformers, are being studied to further improve our model. This would increase the robustness of this approach, which shows promising capabilities for proton range verification.
