Deep reconstruction model for dynamic PET images

Liu H (2017) Deep reconstruction model for dynamic PET images. PLoS ONE 12(9): e0184667. https://doi.org/ 10.1371/journal.pone.0184667 Deep reconstruction model for dynamic PET images Jianan Cui 0 1 Xin Liu 1 Yile Wang 0 1 Huafeng Liu 0 1 ☯ These authors contributed equally to this work. 1 1 0 State Key Laboratory of Modern Optical Instrumentation, Department of Optical Engineering, Zhejiang University , Hangzhou , China , 2 Shenzhen Institutes of Advanced Technology, Chinese Academy of Sciences , Shenzheng , China 1 Editor: Li Zeng, Chongqing University , CHINA Accurate and robust tomographic reconstruction from dynamic positron emission tomography (PET) acquired data is a difficult problem. Conventional methods, such as the maximum likelihood expectation maximization (MLEM) algorithm for reconstructing the activity distribution-based on individual frames, may lead to inaccurate results due to the checkerboard effect and limitation of photon counts. In this paper, we propose a stacked sparse autoencoder based reconstruction framework for dynamic PET imaging. The dynamic reconstruction problem is formulated in a deep learning representation, where the encoding layers extract the prototype features, such as edges, so that, in the decoding layers, the reconstructed results are obtained through a combination of those features. The qualitative and quantitative results of the procedure, including the data based on a Monte Carlo simulation and real patient data demonstrates the effectiveness of our method. - Competing interests: The authors have declared that no competing interests exist. Introduction Positron emission tomography (PET) is a nuclear medicine, functional imaging method that evaluates the body condition based on metabolism in living tissue. Annihilation occurs due to the decay of a radio tracer such as 18F-Fluoro-2-Deoxy-glucose (FDG) after emission positron encounters the electrons, which generates a pair of 511keV photons in the opposite direction. Once the projection data are obtained by a photon detector, the static radioactivity distribution can be reconstructed [ 1 ]. In contrast to static PET imaging, dynamic PET imaging detects data in a series of frames. In clinical applications, the dynamic information has the potential to improve early detection, the characterization of cancer and assessment of therapeutic response after obtaining the spatial and temporal radioactivity distribution [ 2 ]. To guarantee that the rapid change in tracer activity can be tracked immediately, the time interval between two frames in earlier parts of a scan will not be long enough, leading to reduced photon accumulation and lower spatial resolution. How to reconstruct the high-quality dynamic PET images and guarantee the temporal resolution at the same time has become a challenging problem [ 3 ]. One popular method is to reconstruct the radioactivity map for each frame independently based on several conventional statistical models such as maximum a posteriori (MAP) [ 4 ], maximum likelihood estimate (MLE) [ 5 ] and penalized weighted least square (PWLS) [ 6 ]. Because of the tradeoff between spatial resolution and temporal resolution, these frame-toframe strategies may lead to noisy results due to the low signal-to-noise ratio of the data. To solve this problem, an attractive approach is to append the smoothness regularization terms into the objective function. Compared with other approaches such as the filtering of wavelets [ 7 ], total variation (TV) regularization can be suitable for edge-preserving imaging problems in low signal to noise ratio (SNR) or few-view data sets; however, in the presence of noise, TV tends to over smooth and deblur images with a natural gradient [ 8 ]. In this study, we aimed to reduce the noise while preserving the key features (such as the boundary of a tumor) at the same time. Different from regularizations in the above methods, we built a deep reconstruction framework for PET images based on the stacked sparse autoencoder (SAE) and the maximum likelihood expectation maximization (MLEM), which is called MLEM+SAE in short. The SAE model comprised several auto-encoder templates, where each template exploited self-similarity in PET images. To incorporate the information in adjacent frames, temporal features were also learned by the single-layer sparse autoencoder. The essence of our MLEM+SAE model was to incorporate the PET images feature automatically and boost the reconstruction accuracy. Methods The study was approved by the Research Ethics Committee of Zhejiang University. The patients provided written informed consent. Dynamic PET imaging model After injecting the radio tracer into the body, the detected emission data represented the series of photon pairs. The relationship between the detected photon pairs and radioactive distribution could be written as: yiq Poisson…yiq †s:t:yiq ˆ X p Gpqxip …1† …2† where xip is the number of photons emitted from (...truncated)