Development of Novel Image Processing System Using Super-Resolution to Reduce Cone-Beam CT Imaging Dose in Radiation Therapy.
2021
Purpose/Objective(s) To develop an image processing system which can improve the image quality of low-dose Cone-bean CT (CBCT) images for Image Guided Radiation Therapy (IGRT) by using super-resolution technique which can predict high-resolution CBCT images from low-resolution CBCT images. Materials/Methods The developed system was using Single Image Super-Resolution (SISR) based on Convolutional Neural Network (CNN). We used CBCT images of the pelvic region of 123 patients for training of the developed system and its evaluation: 104 patients for training and 19 patients for performance evaluation. Original images (512×512 pixels, IMorg) were downsampled to multiple low matrix sizes (256×256 pixels (IM256), 128×128 pixels (IM128)), and we trained two types of the system to output super-resolution images (512×512 pixels) from each low matrix size image. Then, for performance evaluation, super-resolution images (IMsr2x, IMsr4x) were generated by the trained systems and evaluated them by Peak Signal-to-Noise Ratio (PSNR) and structural similarity (SSIM). Image registration using bone matching and target matching were performed to evaluate the impact of the developed system for image registration. The generated images were also compared with the images obtained by the conventional upsampling method, Bi-Linear (IMlin2x, IMlin4x) and Bi-Cubic (IMcub2x, IMcub4x) methods, using PSNR and SSIM. Results The mean values and standard deviation of PSNR and SSIM of IMsr2x, IMlin2x and IMcub2x were 51.34 ± 0.87 [dB] and 0.995 ± 0.001, 43.05 ± 0.61 [dB] and 0.985 ± 0.002, 45.25 ± 0.66 [dB] and 0.989 ± 0.001, respectively. Similarly, the PSNR and SSIM of IMsr4x, IMlin4x and IMcub4x were 46.27 ± 0.95 [dB] and 0.988 ± 0.002, 37.63 ± 0.61 [dB] and 0.956 ± 0.005, 38.66 ± 0.66 [dB] and 0.960 ± 0.004, respectively. As a result of frequency component analysis, IMsr was added high frequency components that were not included in low matrix size CBCT images. Therefore, the PSNR and SSIM values were significantly higher in IMsr. The registration errors, which were defined as the errors from the registration results of IMorg, of IMsr2x and IMsr4x were 0.20 ± 0.48 [mm] and 1.25 ± 1.45 [mm] for target matching, 0.10 ± 0.31 [mm] and 0.53 ± 0.73 [mm] for bone matching respectively. With IMsr4x, the values of PSNR and SSIM were reduced, but the accuracy of image registration was not significantly affected. Conclusion We developed the image processing system using SISR based on CNN to reduce the CBCT imaging dose in radiation therapy. The developed system could add the high-frequency components which were not included in low matrix size CBCT images. CBCT imaging dose could be reduced to 1/16 and 1/4 when CBCT scans were taken with 128×128 and 256×256 pixels, respectively. The developed system enables CBCT imaging using a low radiation dose without degrading the image quality.
Keywords:
- Correction
- Source
- Cite
- Save
- Machine Reading By IdeaReader
0
References
0
Citations
NaN
KQI