(Sunday, 7/12/2020) [Eastern Time (GMT-4)]
Room: AAPM ePoster Library
Purpose: mitigate this challenge, in this study, a self-supervised framework is proposed for denoising, where only the low-dose CT image is required for training.
Methods: any specified pixel value can be estimated from the rest pixels by exploiting the strong correlations among pixels. This mapping can be learned by masking out one pixel of the input image and then regressing this pixel value in the output. We terms this pixel as the “regularization point”. To incorporate the noisy real observation, another “fidelity point” will be selected for training where the input includes this “fidelity point” and the output tries to learn its pixel value. Specially, to make the training more effective, a set of regularization/fidelity points would be randomly selected chosen from the effective human body region with a CT value larger than -800 HU. But this would induce another problem: the model cannot learn to predict the background information. To alleviate this challenge, a set of “background points” would be randomly selected from the region with CT value less than -800 HU. Overall, the final mean square error loss would be calculated based on the above three different random point sets. The Low-dose CT Challenge 2016 dataset and the popular UNet were used to validate the proposed method. The root mean square error (RMSE) and the structural similarity index (SSIM) were used as the quantitative metrics.
Results: noise in the low-dose input CT image was significantly suppressed while the structures were also well-preserved. Quantitatively speaking, against the normal-dose CT image, the RMSE/SSIM were improved from 88.3/0.64 (input) to 76.0/0.69 (denoised).
Conclusion: CT images can be achieved by only using available low-dose CT images, promoting the deep learning based low-dose CT image denoise algorithm to be more clinically practical.
Keywords
Not Applicable / None Entered.
Taxonomy
Not Applicable / None Entered.
Contact Email
