(Sunday, 7/14/2019) 4:00 PM - 4:30 PM
Room: Exhibit Hall | Forum 9
Purpose: Low Dose Computed Tomography (LDCT) reconstruction aims to obtain high-quality and accurate images, while reducing patient exposure to radiation. This work introduces a block-wised Neural Network toreconstruct CT image with ultra-low dose observations.
Methods: The traditional model-based reconstruction method always formulated as summation of the data term and regularization term, such as regularized least-squares program. Proximal Forward-Backward Splitting (PFBS) method solve the regularized problem by splitting the data term and regularization term and sequentially using gradient descent and proximal method. Inspired by PFBS, we construct a Neural Network with two parts: the network part and the gradient descent part. The aforementioned process could be regarded as a block. While giving an input, we get the output of the block and enforce the distance of output and the ground truth to training the parameters of the network. After obtaining the parameters, we could predict the image and the predicted result is set as the input of a new block. For the whole reconstruction network which contains many blocks, it involves projection and back-projection operations which need auto derivation during backpropagation. In reality, every block is similar to a composition of proximal and gradient descent in PFBS iterations.
Results: We evaluated the proposed neural network with clinical dataset. The neural network is trained with 2D CT images block by block. The proposed method improves the reconstructed image quality in comparison with filtered back-projection (FBP) and TV based method. Furthermore, only a few blocks are available to reconstruct high-quality images.
Conclusion: The block-wised Neural Network method is designed to reconstruct CT from noisy sinogram with low X-ray intensity. Compared with FBP and TV based method, the proposed block-wised Neural Network method suppressed noise in the reconstructed images, while maintaining the structure of organs.
Keywords
Not Applicable / None Entered.
Taxonomy
Not Applicable / None Entered.
Contact Email
