IMPACT OF DIAGNOSTIC MRI IN ADAPTIVE PROTON THERAPY FOR HEAD AND NECK CANCER: AN INNOVATIVE DEEP LEARNING APPROACH FOR CBCT-BASED CT SYNTHESIS

Abstract

Cancer is a pathology at the cellular level where genetic mutation leads to uncontrollable self- replication of cells. Its prevalence and incidence denote the severe effects that the disease pro- duces on patients suffering from it. The pathogenesis associated to cancer is extremely diverse in terms of genes affected, thus, the the treatment selection can become extremely complex.

In the last decades, apart from the technological development of the traditional treatments such as chemotherapy or radiotherapy, a new therapy has gained weight in this field, called proton-therapy. Proton-Therapy is extremely similar to radiotherapy, in terms of physical prin- ciples and mechanisms of action, but the key difference lays in the particles used to irradiate the tumor. Radiotherapy uses photons or electrons, while proton-therapy operates with protons.

Parallel to this development, medical imaging is also thriving, due to the high quality diag- nostic information it provides. In the context of proton-therapy, medical imaging is the main tool used for treatment planning, where three modalities are used. Magnetic Resonance Imaging (MRI), for the initial diagnosis of the tumor, Cone-Beam Computed Tomography (CBCT), for the correct positioning and aligning of the patient, and Computed Tomography (CT) for dose calculations and treatment planning. In this sense, due to the treatment effects, there is anatom- ical tissue variation between weekly treatments. This variation can be stochastic and certain changes are not expected. The CBCT and a control CT are acquired every time the patient has a session scheduled, to assess significant anatomical alterations.

As part of the methodology, in this project, different Deep Learning (DL) architectures have been designed and trained with the objective of synthetically generating the weekly CTs using as input the weekly CBCT and the diagnostic MRI. These architectures comprise Residual U-Nets (ResUnet), Shifted Windows U-Nets (SwinUnet) and Generative Adversarial Networks (GAN). The research outcome was to analyze whether the addition of MRI provides better resolution in soft tissue areas, and to compare the performance of the DL models presented.

For the models¿ results,first qualitative results are presented, and a quantitative analysis was conducted, reporting the MAE, PSNR, SSIM and SPR RE. The model that performed better quantitatively was the ResUnet when having the MRI data as input, and its respective metrics were MAE of 77.529 ± 14.905 HU, PSNR of 27.807 ± 1.775, SSIM of 0.960 ± 0.015 and SPR RE of 5.595 ± 1.010.

With the promising results obtained throughout this project, not only it can be concluded that the MRI aids in improving the soft tissue resolution of the synthetic CT (sCT) generated when performing CBCT-based CT image synthesis, but also that ResUnet models have outperformed the rest of the models quantitatively and qualitatively.