Adaptive proton therapy (APT) removes one of the most significant sources of inaccuracy in treatment delivery, which is using a treatment plan based on an outdated patient anatomy. Adapting the plan throughout the treatment is crucial for delivering an optimal dose to the patient, whose anatomy is constantly changing. This is especially true for proton therapy, where the delivered dose is highly dependent on the range accuracy. Imaging and plan adaptation must be performed online, immediately before the dose delivery, to take maximum advantage of the benefits of APT. The main problem with online APT is that adaptation of the treatment plan takes too long. Therefore, automation of the processes is required to ensure they can be executed adequately in a short time frame.

Deep learning methods have been successfully applied in two processes required for adaptation, namely the definition of structure contours on a CT scan and determining an optimal dose distribution for a given anatomy. Since a treatment plan is dependent on the locations of the different structures, dose prediction methods rely on manually defined contours, which are not available for daily CT scans in APT due to time limitations. This research aims to develop an approach that determines an optimal dose distribution for prostate cancer patients without using manual structure contours.

We use 3D U-Nets for image segmentation and registration as methods for defining the contours on an image. We use another 3D U-Net to predict an optimal dose distribution, which can use predicted or manually defined contours as input. In addition to this, we use two multitask learning approaches that allow one network to perform both contour definition and dose prediction, which makes it possible to share information between the tasks. The first approach is a cross-stitch network that allows two networks to share feature maps if this is beneficial and the second approach is a w-net that consecutively performs contour definition and dose prediction, using the predicted contours for the dose prediction.

The manual contour based dose prediction performed well in the area around the structures, resulting in a test set average 2%/2mm gamma pass rate of 93.4% ± 3.2% and a Dmean prediction error of 0.45% ± 0.36% in the prostate. The average errors for predicting measures such as D95 and V95% in the targets range from 1% to 3%.

The best method for predicting optimal dose distributions without manual contours is to first predict the contours on the CT scan and use those contours for the dose prediction. However, dose predictions based on predicted contours are significantly worse than those based on manual contours, having a 2%/2mm gamma pass rate of 83.8% ± 6.9% and a Dmean prediction error of 0.92% ± 0.7% in the prostate. Their average errors for predicting measures such as D95 and V95% range from 7% to 20%, which makes these predicted dose distributions too inaccurate to be helpful for treatment planning. This shows that dose prediction relies heavily on accurate knowledge of the structure locations, considering the predicted contours have similar quality as those from state-of-the-art methods.

Dose predictions have not improved by additionally learning a network the contour definition task. Using feature maps from other networks via cross-stitch units had no advantageous effect on the predicted dose distributions, mainly because dose predictions not based on structure masks were too bad for it to have any effect. The dose predictions from the w-net did not improve after the segmentation and dose prediction networks were trained together, which could be because the dose prediction loss could not improve the segmentation sufficiently. The main conclusion is that multi-task learning can only benefit related tasks if they can already be performed independently to a certain extent. It is not a substitute for missing information required to perform the task.