Senior Design At Imaging Lab
Information College of Zhejiang University
A reconstruction framework that based on block-dictionary learning and sparse modeling for PET
Author: Yuanyuan Zhu, Huafeng Liu, Nuobei Xie
- Modified BK-SVD and BOMP(dictionary learning) to train block-dictionary of PET data from Sir Run Run Shaw Hosptial
- Presented a reconstruction framework that integrates sparsity penalty on a block-dictionary into Maximum Likelihood estimator which gives a better performance both in image smoothness and accuracy compared to traditional ML-EM method
- Improved the accuracy of PET reconstruction 15% percent than EM-ML
- Abstract:Positron emission tomography (PET) is a kind of radiation tomography, basing on the radioactive tracers which enter into body for tomography. The method for PET image reconstruction is mainly divided into analytical method and iterative methods (statistical method). One of the famous analytical method is called filtered back-projection (FBP) method. The FBP method is based on Radon transform, however, it’s reconstruction accuracy may be low because of the noise in the detection. With the improvement of computer performance, iterative method has become the focus of many researchers. However, iterative method, such as the Maximum Likelihood – Expectation Maximization (ML-EM) method, may have the problem of disease solution. The most a posterior method takes prior knowledge of the image into account, and adds a penalty to ML method, it overcomes the problem of the ML-EM method to a certain extent. Thus, the design of the penalty item becomes the focus. The reconstruction method based on sparse representation and dictionary learning can be designed as a penalty term, because they can use the prior knowledge of some anatomical images such as CT and MRI. On the other hand, block dictionary learning can achieve smaller representation error rates than ordinary dictionary learning method. Therefore, this paper presents a PET image reconstruction method based on block dictionary learning and sparse representation. We use block dictionary learning method BK-SVD to train a dictionary and its sparse representation as the penalty items (a prior function) and use a likelihood function based on Poisson’s distribution. The prior function with the likelihood function forms a maximum a posteriori estimation problem.
In order to verify the proposed reconstruction model, this paper performs the image reconstruction experiments based on the Monte Carlo simulation data of chest and brain ,
comparing with the ML-EM method and the SPS-OS method. We also analyze the influence of different parameters of the model.
Reconstruction from PET data
With presented framework, improved the accuracy of reconstruction
14 x 14 BK dictionary
Trained with BKSVD and BOMP from a CT image.