Program Information
Dose Reduction in X-Ray Computed Tomography Via a Deep-Learning Approach
Y Gonzalez*, C Shen , B Li , P Klages , X Jia , The University of Texas Southwestern Medical Center, Dallas, TX
Presentations
SU-K-FS4-13 (Sunday, July 30, 2017) 4:00 PM - 6:00 PM Room: Four Seasons 4
Purpose: As the use of CT continues to increase in popularity, radiation dose to the patient becomes a concern. It is desirable to develop effective methods to reduce noise level in CT to allow for radiation dose reduction. In this study, we propose to use a deep-learning approach to denoise CT projection images.
Methods: We used 12 CT images of different patients and anatomical sites. Noise-free and noisy CT projections were generated for these CT images. The projection images were subdivided into patches with a size of 16 by 16 pixels. Patches from nine noise-free and corresponding noisy projections were input to a multi-layer convolutional neural network (CNN) to train a denoising model that is capable of reducing noise from an input noisy patch. After the training stage, we applied the model to every patch in a testing CT projection image to reduce noise level. The denoised patches were then assembled into a complete denoised projection. Finally, a CT image was reconstructed using the denoised projection image. To evaluate this approach, we compared results with those from BM3D and median filtering denoising methods.
Results: Structural features in the CT images generated by the CNN method were sharper than those in the images processed by the BM3D method. Quantitatively, the CNN method was able to outperform both the median filter and BM3D methods in all cases. On average, the CNN method reduced the CT number error from 29% in the unprocessed case to 11%, in comparison to the error of 14% in the BM3D method.
Conclusion: The proposed CNN-based denoising method was found to be effective in terms of reducing noise level in a CT image, which may allow for radiation dose reduction in CT.
Contact Email: