Systems and methods for image denoising via adversarial learning
View Patent ↗Various examples are provided related to reconstructing images such as, e.g., medical images from low-dose image scans. Adversarial learning such as, e.g., a Cyclic Simulation and Denoising (CSD) framework can be used to address challenges of complicated mixed noise in real low-dose scans. The CSD framework can include a simulator model that can extract low-dose noise and features (e.g., tissue features) from separate image spaces into a unified feature space and a denoiser model that can learn how to remove noise and restore features, simultaneously. Both the simulator model and the denoiser model can regularize each other in a cyclic manner to optimize network learning effectively. The CSD framework in combination with phantom scans can embrace the realistic low-dose noise and features into a unified learning environment to address the challenge of real low-dose image restoration.
1 . An adversarial machine learning and denoising system, comprising:
a memory device storing a simulator model and a denoiser model; and
a processor configured to perform the simulator model and the denoiser model, wherein during performance of a simulation-to-denoising (S2D) training cycle of the simulator model and the denoiser model, the simulator model receives as input a low-dose noisy phantom image scan and a high-dose patient image scan and uses the low-dose noisy phantom image scan and the high-dose patient image scan to generate a simulated low-dose noisy patient image scan, and the denoiser model receives the simulated low-dose noisy patient image scan output from the simulator model and uses the simulated low-dose noisy patient image scan to train the denoiser model to remove noise from a real low-dose noisy patient image scan, where low-dose image scans are performed at about 95 mAs or less and high-dose image scans are performed above the low-dose image scans.
2 . The adversarial machine learning and denoising system of claim 1 , wherein the denoiser model operates as a regularizer for the simulator model during the S2D training cycle by outputting feedback to the simulator model characterizing a noise level of the simulated low-dose noisy patient image scan output by the simulator model, the simulator model using the feedback to train the simulator model to improve the noise level of the simulated low-dose noisy patient image scan.
3 . The adversarial machine learning and denoising system of claim 2 , wherein during performance of a denoising-to-simulation (D2S) training cycle of the simulator model and the denoiser model, the denoiser model receives as input a low-dose noisy patient image scan and a low-dose noisy phantom image scan and uses the low-dose noisy patient image scan and the low-dose noisy phantom image scan to generate a generated high-dose patient image scan, and the simulator model receives the generated high-dose patient image scan output from the denoiser model and uses the generated high-dose patient image scan and the low-dose noisy phantom image scan to train the simulator model to generate the simulated low-dose noisy patient image scan.
4 . The adversarial machine learning and denoising system of claim 3 , wherein the simulator model operates as a regularizer for the denoiser model during the D2S training cycle by outputting feedback to the denoiser model that the denoiser model uses to train the denoiser model to improve a noise level of the generated high-dose patient image scan output by the denoiser model.
5 . The adversarial machine learning and denoising system of claim 1 , wherein phantom image scans are obtained with an anthropomorphic physical phantom model.
6 . An adversarial learning and denoising method, comprising:
during performance of a simulation-to-denoising (S2D) training cycle, receiving as input in a simulator model a low-dose noisy phantom image scan and a high-dose patient image scan and outputting a simulated low-dose noisy patient image scan, where low-dose image scans are performed at about 95 mAs or less and high-dose image scans are performed above the low-dose image scans; and
during performance of the S2D training cycle, receiving the simulated low-dose noisy patient image scan output from the simulator model in a denoiser model and using the simulated low-dose noisy patient image scan in the denoiser model to train the denoiser model to remove noise from a real low-dose noisy patient image scan.
7 . The adversarial learning and denoising method of claim 6 , further comprising:
during the S2D training cycle, operating the denoiser model as a regularizer for the simulator model by outputting feedback to the simulator model characterizing a noise level of the simulated low-dose noisy patient image scan output by the simulator model; and
during the S2D training cycle, using the feedback in the simulator model to train the simulator model to improve the noise level of the simulated low-dose noisy patient image scan.
8 . The adversarial learning and denoising method of claim 7 , further comprising:
during performance of a denoising-to-simulation (D2S) training cycle, receiving as input in the denoiser model a low-dose noisy patient image scan and a low-dose noisy phantom image scan and using the low-dose noisy patient image scan and the low-dose noisy phantom image scan to remove noise from the low-dose noisy patient image scan to generate a generated high-dose patient image scan; and
during performance of the D2S training cycle, receiving the generated high-dose patient image scan output from the denoiser model in the simulator model and using the generated high-dose patient image scan to train the simulator model to generate the low-dose noisy patient image scan.
9 . The adversarial learning and denoising method of claim 8 , further comprising:
during the D2S training cycle, operating the simulator model as a regularizer for the denoiser model by outputting feedback to the denoiser model.
10 . The adversarial learning and denoising method of claim 9 , further comprising:
during the D2S training cycle, using the feedback output by the simulator model in the denoiser model to train the denoiser model to improve the noise level of the generated high-dose patient image scan generated by the denoiser model.
11 . The adversarial learning and denoising method of claim 6 , wherein phantom image scans are obtained with an anthropomorphic physical phantom model.