(Sunday, 7/12/2020) [Eastern Time (GMT-4)]
Room: AAPM ePoster Library
Purpose: Model-based iterative reconstruction (MBIR) is popular for low-dose CT (LDCT) for its flexibility to model data and image priors. However, the choice of appropriate hyper-parameters in MBIR is often manually done and may not be optimized. This work will develop deep learning (DL) based MBIR to automatically optimize the hyper-parameters and enable the learning with different low-dose levels.
Methods: In our new DL-regularized MBIR method, image priors and hyper-parameters are collaboratively optimized by two neural networks (NN). The proposed method is based on the unrolling of so-called half-quadratic splitting method, and each stage contains two blocks: (1) inversion block reconstructs image using the measurements and the previous estimate, whose hyper-parameters are predicted by a fully connected network (FCN); (2) de-noising block removes artifacts of the estimate passed from inverse block, by a convolution neural network (CNN). Firstly, the regularization used in the inversion block is based on fine-grained high-pass filter banks motivated from wavelet tight frames. Secondly, the inner-loop hyper-parameters for regularization are predicted by a FCN learned from the residual.
Results: The performance of the proposed methods is evaluated in comparison with FBP method, TV-based PWLS method and FBPConvNet (a DL-based image postprocessing method). Experimental evaluation on clinical patient dataset showed that the proposed method provided the best image reconstruction quality in terms of both quantitative metrics and visual assessments.
Conclusions: We have proposed an image reconstruction method for LDCT driven by a dual pair of NN. One is for denoising intermediate results, and the other NN is for predicting optimal hyper-parameters involved in image reconstruction. Different from existing approaches, the hyper-parameters involved in our approach is adaptive to the measurements obtained with different X-ray intensity and adaptive to different target images. The model trained by the proposed method is applicable to CT imaging with different dose levels.
Keywords
Not Applicable / None Entered.
Taxonomy
Not Applicable / None Entered.
Contact Email
