Deep learning-based metal artifact reduction using cycle-consistent adversarial network for intensity-modulated head and neck radiation therapy treatment planning

Abstract

To develop a deep learning-based metal artifact reduction (DL-MAR) method using unpaired data and to evaluate its dosimetric impact in head and neck intensity-modulated radiation therapy (IMRT) compared with the water density override method. The data set comprised the data of 107 patients who underwent radiotherapy. Fifteen patients with dental fillings were used as the test data set. The computed tomography (CT) images of the remaining 92 patients were divided into two domains: the metal artifact and artifact-free domains. CycleGAN was used for domain translation. The artifact index of the DL-MAR images was compared with that of the original uncorrected (UC) CT images. The dose distributions of the DL-MAR and UC plans were created by comparing the reference clinical plan with the water density override method (water plan) in each dataset. Dosimetric deviation in the oral cavity from the water plan was evaluated. The artifact index of the DL-MAR images was significantly smaller than that of the UC images in all patients (13.2±4.3 vs. 267.3±113.7). Compared with the reference water plan, dose differences of the UC plans were greater than those of the DL-MAR plans. DL-MAR images provided dosimetric results that were more similar to those of the water plan than the UC plan. We developed a fast DL-MAR method using CycleGAN for head and neck IMRT. The proposed method could provide consistent dose calculation against metal artifact and improve the efficiency of the planning process by eliminating manual delineation.