Room: Room 202
Purpose: Metal implants cause artifacts in CT images, which adversely affect accuracy of diagnosis and radiotherapy. Metal artifact reduction (MAR) methods usually approach the problem by restoring projection data corrupted by metals via use of handcrafted priors, such as linear interpolation (LI) or normalization (NMAR). Recent advancements in deep learning have shown unique and effective ways to incorporate prior knowledge for superior image recognition and reconstruction. In this study, we propose an MAR method to restore CT sinograms with a deep convolutional image generator.
Methods: First, projection data tracing metal is distinguished in the sinogram. Then, tissue segmentation based on a coarse MAR result is used to generate a prior sinogram. Next, metal regions of the difference image between measured and prior sinograms is restored by a deep image prior-based convolutional decoder-encoder network (DIP). Finally, the conventional reconstruction is performed on the synthetic sinogram. One simulation and one real dataset were used to evaluate the proposed method. In simulation, two virtual metallic traces in the sinogram of the Shepp-Logan phantom was assumed as needed to be restored. For real data, a CIRS phantom with two metallic inserts was scanned on CBCT.
Results: From results using both simulation and real data, the proposed method can be seen to reduce metal artifacts effectively. Specifically, in the simulation, RMSE of the reconstruction is only 0.0013 for the proposed DIP method, where it is 0.0036 and 0.0083 for LI and NMAR methods, respectively. In the real data experiment, the standard deviation of the water region is 0.0105/cm for DIP, where it is 0.0109/cm, 0.0119/cm, and 0.0211/cm for LI, NMAR, and reconstruction without MAR, respectively.
Conclusion: By using the proposed sinogram handling strategy and a deep image prior-based inpainting technique, the projection data through metals can be restored very well, leading to satisfactory image reconstruction.