Kentaro Miki1, Martijn Kusters2, Takeo Nakashima3, Akito Saito4, Daisuke Kawahara5, Ikuno Nishibuchi6, Tomoki Kimura4, Yuji Murakami6, Yasushi Nagata6. 1. Department of Radiation Oncology, Hiroshima University Hospital, Hiroshima, Japan. Electronic address: kentaro-miki@hiroshima-u.ac.jp. 2. Department of Radiation Oncology, Radboud University Medical Center, Nijmegen, the Netherlands. 3. Radiation Therapy Section, Department of Clinical Support, Hiroshima University Hospital, Hiroshima, Japan. 4. Department of Radiation Oncology, Hiroshima University Hospital, Hiroshima, Japan. 5. Department of Radiation Oncology, Graduate School of Biomedical & Health Sciences, Hiroshima University, Hiroshima, Japan. 6. Department of Radiation Oncology, Hiroshima University Hospital, Hiroshima, Japan; Department of Radiation Oncology, Graduate School of Biomedical & Health Sciences, Hiroshima University, Hiroshima, Japan.
Abstract
PURPOSE: Lack of a reference dose distribution is one of the challenges in the treatment planning used in volumetric modulated arc therapy because numerous manual processes result from variations in the location and size of a tumor in different cases. In this study, a predicted dose distribution was generated using two independent methods. Treatment planning using the predicted distribution was compared with the clinical value, and its efficacy was evaluated. METHODS: Computed tomography scans of 81 patients with oropharynx or hypopharynx tumors were acquired retrospectively. The predicted dose distributions were determined using a modified filtered back projection (mFBP) and a hierarchically densely connected U-net (HD-Unet). Optimization parameters were extracted from the predicted distribution, and the optimized dose distribution was obtained using a commercial treatment planning system. RESULTS: In the test data from ten patients, significant differences between the mFBP and clinical plan were observed for the maximum dose of the brain stem, spinal cord, and mean dose of the larynx. A significant difference between the dose distributions from the HD-Unet dose and the clinical plan was observed for the mean dose of the left parotid gland. In both cases, the equivalent coverage and flatness of the clinical plan were observed for the tumor target. CONCLUSIONS: The predicted dose distribution was generated using two approaches. In the case of the mFBP approach, no prior learning, such as deep learning, is required; therefore, the accuracy and efficiency of treatment planning will be improved even for sites where sufficient training data are unavailable.
PURPOSE: Lack of a reference dose distribution is one of the challenges in the treatment planning used in volumetric modulated arc therapy because numerous manual processes result from variations in the location and size of a tumor in different cases. In this study, a predicted dose distribution was generated using two independent methods. Treatment planning using the predicted distribution was compared with the clinical value, and its efficacy was evaluated. METHODS: Computed tomography scans of 81 patients with oropharynx or hypopharynx tumors were acquired retrospectively. The predicted dose distributions were determined using a modified filtered back projection (mFBP) and a hierarchically densely connected U-net (HD-Unet). Optimization parameters were extracted from the predicted distribution, and the optimized dose distribution was obtained using a commercial treatment planning system. RESULTS: In the test data from ten patients, significant differences between the mFBP and clinical plan were observed for the maximum dose of the brain stem, spinal cord, and mean dose of the larynx. A significant difference between the dose distributions from the HD-Unet dose and the clinical plan was observed for the mean dose of the left parotid gland. In both cases, the equivalent coverage and flatness of the clinical plan were observed for the tumor target. CONCLUSIONS: The predicted dose distribution was generated using two approaches. In the case of the mFBP approach, no prior learning, such as deep learning, is required; therefore, the accuracy and efficiency of treatment planning will be improved even for sites where sufficient training data are unavailable.
Authors: Martijn Kusters; Kentaro Miki; Liza Bouwmans; Karl Bzdusek; Peter van Kollenburg; Robert Jan Smeenk; René Monshouwer; Yasushi Nagata Journal: Phys Imaging Radiat Oncol Date: 2022-02-03