(Sunday, 7/29/2018) 5:05 PM - 6:00 PM
Room: Room 207
Purpose: To remove noise and scatter of low dose Cone Beam CT (CBCT) images using deep learning convolutional neural network (CNN) for patient setup in radiation oncology treatment.
Methods: Both phantom and patient data are used in this study. A digital anthropomorphic phantom was used to create simulation CBCT projections with or without random scatter and noise. A CNN network structure with linear and non-linear filters layer was built to iteratively reduce the scatter and noise. Various architectures of the CNN has been tested and adjusted to optimize its performance. Thereafter, the product CNN network was applied to a low dose patients setup CBCT projection image set with noise and scatter. Projections generated from a fan beam CT image was used as reference image which was considered as projections with minimal noise and scatter. Reference-based image quality metrics, root mean square error (RMSE) and peak signal to noise ratio (PSNR), are employed to comprehensively evaluate the performance of the architecture at improving image quality.
Results: The final CNN network structure can automatically optimize it’s parameters to adaptively reduce the impact of noise and scatter on CBCT projections. The root mean square error (RMSE) of 0.0256 and peak signal to noise ratio (PSNR) of 80.0467 indicates an improvement in image quality from the original, noisy CBCT scans which is 0.0647(RMSE) and 71.9603(PSNR)for patient data. Comparing with other imaging processing methods, the CNN can provide an adaptive noise and scatter reduction and better improvement of image quality with p-value less than 0.01.
Conclusion: The application of deep leaning CNN will significantly improve the quality of CBCT images without increasing radiation dose to the patient.
Keywords
Not Applicable / None Entered.
Taxonomy
Not Applicable / None Entered.
Contact Email
