Room: Exhibit Hall | Forum 1
Purpose: Low dose CT by reducing either X-ray exposure (mAs) or the number of X-ray projections will elevate image noise level or result in obvious streaking artifacts if using filtered backprojection (FBP) image reconstruction algorithm. In this study, we propose a novel deep-learning based method to reconstruct low dose CT simulation images to reduce image noise and mitigate artifacts.
Methods: The proposed method integrates a residual block concept into a cycle-consistent adversarial network framework to learn a mapping between low dose and full dose CT images. A fully convolution neural network with residual blocks is used in the generator to enable end-to-end transformations. The full dose CT images of 17 brain patients were used to simulate corresponding low-projection or low-mAs CT images contaminated with noise and artifacts. These low dose images were then fed into our network to retrieve high quality CT images. The original full dose images were considered as ground truth for evaluation purpose
Results: The image noise and artifacts were highly suppressed on the low-mAs and low-projection CT images generated by our method. With the mAs level reduced to 0.5% of that of full dose, the average mean absolute error (MAE) and image signal to noise ratio (SNR) of the images generated by our method are 25.3Â±11.0 HU and 0.88Â±0.40, compared with 56.3Â±11.9 HU and 0.45Â±0.13 of FBP results. With the number of projection decreased to 60 (1/16 of full projections), the average MAE of the images obtained by our method are 18.3Â±2.5 HU compared with 63.0Â±2.8 HU of FBP results.
Conclusion: We have developed a novel deep learning-based approach to effectively suppress image noise and streaking artifacts for low dose CT. The efficacy of our method has been well demonstrated by the significantly improved image quality compared with conventional FBP reconstruction method.