Room: AAPM ePoster Library
Purpose: aim of this study was to develop the novel image processing system that can reduce CT exposure dose while improving image quality by super-resolution using deep learning.
Methods: developed system using Convolutional Neural Network (ResNet) enabled to produce high matrix size images from low matrix size images. Original images (IM(org)) with 512 by 512 pixels were downsampled to multiple low matrix sizes (128 by 128 pixels (IM128), 256 by 256 pixels (IM256)), and then super-resolution images with 512 by 512 pixels (IM(sr)) were predicted from each matrix size and compared with the images produced by bi-cubic method (IM(cubic)). CT images of 57 prostate cancer patients were used to train the developed system and the images of 5 patients were used for performance evaluation. Evaluation of IM(sr) was implemented using Peak Signal-to-Noise Ratio (PSNR) and Structural Similarity (SSIM).
Results: the case of super-resolution of IM128, the mean and standard deviation values of PSNR and SSIM of IM(cubic) and IM(sr) were 34.85 ± 0.39[dB], 0.945 ± 0.004 and 46.95 ± 1.34 [dB], 0.993 ± 0.001, respectively. In the case of super-resolution of IM256, the mean and standard deviation values of PSNR and SSIM of IM(cubic) and IM(sr) were 43.09 ± 0.51[dB], 0.991 ± 0.001 and 53.99 ± 1.12 [dB], 0.998 ± 0.001, respectively. In addition, high-frequency components lost due to dawnsampling were restored in IMsr.
Conclusion: developed the novel image processing system which can reduce CT exposure dose with improving image quality. Exposure dose could be reduced to 1/16 and 1/4 when CT scans were taken with 128 by 128 pixels and 256 by 256, respectively. With the developed system, it was possible to predict high matrix size images with high frequency components restored from low matrix size images taken at low dose.