Conformal energy modulator design optimization using deep learning for FLASH proton therapy treatment planning

Abstract

Radiotherapy is a method used to treat cancers based on the utilization of photons to target and destroy cancer cells. One of the radiotherapy techniques called proton ther- apy uses the energy released by protons to kill cancer cells. Compared to conventional radiotherapy that uses photons, proton therapy is known to spare more healthy tissues surrounding the tumor. This advantage stems from the particular dose (amount of energy absorbed per kg) profile of protons in matter called the Bragg peak. When proton therapy is delivered at very high dose rates (>40 Gy/s), this technique is called FLASH proton therapy. It has the advantage of sparing even more healthy tissues. The extra short delivery duration of FLASH proton therapy makes proton adminis- tration technically challenging and some delivery methods must be adapted. Importantly, the technique to control the depth dose distribution within the tumor must be entirely rethought. To address this, a new device must be added to the trajectory of the proton beams. This new piece named a Conformal Energy Modulator, is made of a square base on top of which spikes of different heights rise. Its configuration must be tuned accord- ing to the tumor shape. The CEM optimization can be achieved by treatment planning systems but current solutions either lack accuracy or take too long to compute. In this master’s thesis, a new method to optimize the CEM quickly and accurately is proposed, relying on an artificial intelligence model called Convolutional Neural Network. This thesis is divided into 5 main parts. The first part introduces how proton beams are generated and delivered to the patients, how protons interact with matter, and why their dose profile is so interesting from a treatment point of view. Additionally, the chal- lenges posed by the FLASH concept and the required modifications for proton therapy administration to accommodate FLASH PT are discussed. The second part is related to the data acquisition needed for the AI model training. It is performed with a treatment planning system that simulates FLASH PT with the CEM and the resulting dose dis- tribution. The CEMs profiles will serve as outputs and their corresponding dose maps will serve as inputs of the AI model. The third part introduces the concepts of artificial intelligence, neural networks, and convolutional neural networks. The choice of the spe- cific AI model to solve this problem is detailed as well as its architecture. The fourth part discusses the training of the model. The model’s ability to predict a CEM given a dose map is evaluated using validation and test sets. Lastly, the performance of the predicted CEMs by the model compared to the initial CEMs of the data set is evalu- ated through a comparison of their dose maps resulting from FLASH PT simulations. The dose maps comparison is performed with dose-volume histograms and Gamma index evaluation. This comparison highlights the significant similarities observed in the dose maps, indicating the model’s excellent training and successful CEM optimization.