REGISTRATION APPROACH FOR CBCT BASED CT SYNTHESIS IN ADAPTIVE PROTON THERAPY FOR HEAD AND NECK CANCER

Abstract

Cancer is a cellular pathology characterized by genetic mutations that lead to uncontrolled cell replication. Its prevalence and impact highlight the significant effects on patients. The pathogenesis of cancer varies widely based on the affected genes, making treatment selection complex. In recent decades, alongside advancements in traditional treatments like chemother apy and radiotherapy, proton therapy has emerged as a prominent new therapy. While proton therapy is similar to radiotherapy in terms of physical principles and mechanisms of action, it differs in the particles used for tumor irradiation. Radiotherapy uses photons or electrons, whereas proton therapy uses protons. Concurrent with these developments, medical imaging has seen significant advancements due to its provision of high-quality diagnostic information. In the context of proton therapy, medical imaging plays a crucial role in treatment planning, utilizing Cone-Beam Computed To mography (CBCT) for patient positioning and alignment, and Computed Tomography (CT) for dose calculations and treatment planning. Due to treatment effects, anatomical tissue variations can occur between weekly sessions. CBCT scans are employ to control and assess anatomical changes during treatment. With regards to the methodology, we conducted a retrospective study with sCT images generated from CBCT scans and CT HU scales (product of previous deep learning research) frompCTusedintheoriginalplanningofthepatients¿ treatment. Weekly CTs acquired between treatment sessions (necessary for recalculating doses due to anatomical changes), were used as ground truth for comparison of 3 different registration methodologies to leverage sCT volumes and account for field-of-view limitations. These images were collected for 7 patients and 26 sCT-pCT pairs with General Electric Revolution CT scanners. Patients had underwent proton therapy treatment at Quironsalud Proton-Therapy Center. Qualitative analysis of the models is performed through visual inspection and qualitative analysis is done by reporting similarity metrics between the resulting volumes and the ground truth. The MeanSquaredErrorvalueswere0.006959fortheNon-Deformablemethod, 0.009312 for Elastix, and 0.002932 for VoxelMorph. The Mean Absolute Error (MAE) values were 2.472341×10¿7 for the Non-Deformable method, 3.098558×10¿7 for Elastix, and 7.441205× 10¿7 for VoxelMorph. The Structural Similarity Index (SSI) values were 0.727512 for the Non Deformable method, 0.710489 for Elastix, and 0.770104 for VoxelMorph. The Mutual Infor mation (MI) values were 0.532818 for the Non-Deformable method, 0.531219 for Elastix, and 0.815192 for VoxelMorph. Judging by the metric summary VoxelMorph methodology seemed to performed better than the tailored non-deformable and elastic methods. However it could not be conclud statistically which method was better