Super-resolution of dose distributions from a two-dimensional array detector using a convolutional neural network

다국어 초록 (Multilingual Abstract)

Gamma index pass rate is a commonly used to verify the accuracy of complex radiation therapy. To calculate the accurate gamma index pass rate, a pixel spacing of less than 1/3 of the criteria distance is required. Film is providing fne resolution for measuring dose distribution several disadvantages have led to a preference for 2D array detectors. However, 2D array detectors have lower spatial resolution, which requires improvement to calculate the accurate gamma pass rate. Interpolation methods have traditionally been used for this purpose. While some researchers have made eforts to increase the acquired spatial resolution, the limitation of having to fx the gantry at 0 degrees for irradiation. To overcome this limitation, this study applies super-resolution using convolutional neural networks (CNN) to dose distribution acquisition. To train the model on a typical ofce computer, the efcient sub-pixel convolutional neural network (ESPCN) was used, which has excellent peak signal-to-noise ratio (PSNR) and structural similarity index (SSIM) compared to the number of parameters required. To train the model, the required dose distribution was obtained from the publicly available Non-Small Cell Lung Cancer (NSCLC) Radiomics data set contained in The Cancer Image Archive (TCIA). This data set consists of computed tomography (CT) images and the radiotherapy structure set (RTSTRUCT), which contains location information for tumours and critical organs.
To evaluate the performance of the trained model, the PSNR and SSIM were compared for dose distributions obtained from the treatment planning system (TPS), those with super-resolution applied using ESPCN, and those with bi-linear and bi-cubic interpolation. The PSNR values for ESPCN, bi-linear and bi-cubic were 36.34, 32.57 and 34.84, respectively, while the corresponding SSIM values were 0.9496, 0.9470 and 0.9566. In addition, 97.37%, 90.96% and 94.05% gamma pass rates were measured under the 3%/2 mm condition. When comparing various evaluation metrics, it was concluded that the dose distribution with super-resolution was superior to the case with interpolation.

참고문헌 (Reference)

• 1 D. A. Low, 25 : 656-, 1998
• 2 M. Miften, 45 : e53-, 2018
• 3 M. Wendling, 34 : 1647-, 2007
• 4 M. Chun, 18 : 521-, 2020
• 5 J. Son, 49 : 307-, 2015
• 6 E. Spezi, 31 : 269-, 2006
• 7 T. S. Stelljes, 42 : 1528-, 2015
• 8 A. S. Oinam, 34 : 153-, 2009
• 9 S. Stathakis, 20 : 1154-, 2015
• 10 M. Hussein, 36 : 1-, 2017
• 11 J. G. Li, 10 : 62-74, 2009
• 12 S. Ko, 22 : 33-, 2010
• 13 K. Nasrollahi, 25 : 1423-, 2014
• 14 J. Zhou, 2347 : 012021-, 2022
• 15 C. Dong, 38 : 295-, 2015
• 16 W. Shi, 1874-1883, 2016
• 17 J. Kim, 1646-1654, 2016
• 18 B. Lim, 136-144, 2017
• 19 H. J. Aerts, 5 : 4006-, 2014
• 20 K. Clark, 26 : 1045-, 2013
• 21 K. Mei, 13 : 591-, 2019
• 22 "The jamovi project, computer software"
• 23 전호상 ; Ji Hyeon Joo ; 기용간 ; 김원택 ; 남지호 ; 김동현 ; Dahl Park, "A Method to Enhance the Spatial Resolution of the Two-dimensional Detector Arrays for the Precise Dose Assessment of Intensity Modulated Radiation Therapy" 한국물리학회 75 (75): 1048-1053, 2019