Iori Sumida1, Taiki Magome2, Indra J Das3, Hajime Yamaguchi4, Hisao Kizaki4, Keiko Aboshi4, Hiroko Yamaguchi4, Yuji Seo1, Fumiaki Isohashi1, Kazuhiko Ogawa1. 1. Department of Radiation Oncology, Osaka University Graduate School of Medicine, 2-2 Yamada-oka, Suita, Osaka 565-0871 Japan. 2. Department of Radiological Sciences, Faculty of Health Sciences, Komazawa University, 1-23-1 Komazawa, Setagaya-ku, Tokyo 154-8525 Japan. 3. Department of Radiation Oncology, Northwestern Memorial Hospital, Northwest University Medical Center, Galter Pavilion, Chicago, IL 60611. 4. Department of Radiation Oncology, Daini Osaka Police Hospital, 2-6-40 Karasugatsuji, Tennoji-ku, Osaka 543-8922 Japan.
Abstract
PURPOSE: This study aims to investigate the feasibility of using convolutional neural networks to predict an accurate and high resolution dose distribution from an approximated and low resolution input dose. METHODS: Sixty-six patients were treated for prostate cancer with VMAT. We created the treatment plans using the Acuros XB algorithm with 2 mm grid size, followed by the dose calculated using the anisotropic analytical algorithm with 5 mm grid with the same plan parameters. U-net model was used to predict 2 mm grid dose from 5 mm grid dose. We investigated the two models differing for the training data used as input, one used just the low resolution dose (D model) and the other combined the low resolution dose with CT data (DC model). Dice similarity coefficient (DSC) was calculated to ascertain how well the shape of the dose-volume is matched. We conducted gamma analysis for the following: DVH from the two models and the reference DVH for all prostate structures. RESULTS: The DSC values in the DC model were significantly higher than those in the D model (p < 0.01). For the CTV, PTV, and bladder, the gamma passing rates in the DC model were significantly higher than those in the D model (p < 0.002-0.02). The mean doses in the CTV and PTV for the DC model were significantly better matched to those in the reference dose (p < 0.0001). CONCLUSIONS: The proposed U-net model with dose and CT image used as input predicted more accurate dose.
PURPOSE: This study aims to investigate the feasibility of using convolutional neural networks to predict an accurate and high resolution dose distribution from an approximated and low resolution input dose. METHODS: Sixty-six patients were treated for prostate cancer with VMAT. We created the treatment plans using the Acuros XB algorithm with 2 mm grid size, followed by the dose calculated using the anisotropic analytical algorithm with 5 mm grid with the same plan parameters. U-net model was used to predict 2 mm grid dose from 5 mm grid dose. We investigated the two models differing for the training data used as input, one used just the low resolution dose (D model) and the other combined the low resolution dose with CT data (DC model). Dice similarity coefficient (DSC) was calculated to ascertain how well the shape of the dose-volume is matched. We conducted gamma analysis for the following: DVH from the two models and the reference DVH for all prostate structures. RESULTS: The DSC values in the DC model were significantly higher than those in the D model (p < 0.01). For the CTV, PTV, and bladder, the gamma passing rates in the DC model were significantly higher than those in the D model (p < 0.002-0.02). The mean doses in the CTV and PTV for the DC model were significantly better matched to those in the reference dose (p < 0.0001). CONCLUSIONS: The proposed U-net model with dose and CT image used as input predicted more accurate dose.