PET scatter estimation using deep learning U-Net architecture.

Objective.Positron emission tomography (PET) image reconstruction needs to be corrected for scatter in order to produce quantitatively accurate images. Scatter correction is traditionally achieved by incorporating an estimated scatter sinogram into the forward model during image reconstruction. Existing scatter estimated methods compromise between accuracy and computing time. Nowadays scatter estimation is routinely performed using single scatter simulation (SSS), which does not accurately model multiple scatter and scatter from outside the field-of-view, leading to reduced qualitative and quantitative PET reconstructed image accuracy. On the other side, Monte-Carlo (MC) methods provide a high precision, but are computationally expensive and time-consuming, even with recent progress in MC acceleration.Approach.In this work we explore the potential of deep learning (DL) for accurate scatter correction in PET imaging, accounting for all scatter coincidences. We propose a network based on a U-Net convolutional neural network architecture with 5 convolutional layers. The network takes as input the emission and computed tomography (CT)-derived attenuation factor (AF) sinograms and returns the estimated scatter sinogram. The network training was performed using MC simulated PET datasets. Multiple anthropomorphic extended cardiac-torso phantoms of two different regions (lung and pelvis) were created, considering three different body sizes and different levels of statistics. In addition, two patient datasets were used to assess the performance of the method in clinical practice.Main results.Our experiments showed that the accuracy of our method, namely DL-based scatter estimation (DLSE), was independent of the anatomical region (lungs or pelvis). They also showed that the DLSE-corrected images were similar to that reconstructed from scatter-free data and more accurate than SSS-corrected images.Significance.The proposed method is able to estimate scatter sinograms from emission and attenuation data. It has shown a better accuracy than the SSS, while being faster than MC scatter estimation methods.

CASToR: a generic data organization and processing code framework for multi-modal and multi-dimensional tomographic reconstruction.

In tomographic medical imaging (PET, SPECT, CT), differences in data acquisition and organization are a major hurdle for the development of tomographic reconstruction software. The implementation of a given reconstruction algorithm is usually limited to a specific set of conditions, depending on the modality, the purpose of the study, the input data, or on the characteristics of the reconstruction algorithm itself. It causes restricted or limited use of algorithms, differences in implementation, code duplication, impractical code development, and difficulties for comparing different methods. This work attempts to address these issues by proposing a unified and generic code framework for formatting, processing and reconstructing acquired multi-modal and multi-dimensional data. The proposed iterative framework processes in the same way elements from list-mode (i.e. events) and histogrammed (i.e. sinogram or other bins) data sets. Each element is processed separately, which opens the way for highly parallel execution. A unique iterative algorithm engine makes use of generic core components corresponding to the main parts of the reconstruction process. Features that are specific to different modalities and algorithms are embedded into specific components inheriting from the generic abstract components. Temporal dimensions are taken into account in the core architecture. The framework is implemented in an open-source C++ parallel platform, called CASToR (customizable and advanced software for tomographic reconstruction). Performance assessments show that the time loss due to genericity remains acceptable, being one order of magnitude slower compared to a manufacturer's software optimized for computational efficiency for a given system geometry. Specific optimizations were made possible by the underlying data set organization and processing and allowed for an average speed-up factor ranging from 1.54 to 3.07 when compared to more conventional implementations. Using parallel programming, an almost linear speed-up increase (factor of 0.85 times number of cores) was obtained in a realistic clinical PET setting. In conclusion, the proposed framework offers a substantial flexibility for the integration of new reconstruction algorithms while maintaining computation efficiency.

Dynamic PET image reconstruction integrating temporal regularization associated with respiratory motion correction for applications in oncology.

Respiratory motion reduces both the qualitative and quantitative accuracy of PET images in oncology. This impact is more significant for quantitative applications based on kinetic modeling, where dynamic acquisitions are associated with limited statistics due to the necessity of enhanced temporal resolution. The aim of this study is to address these drawbacks, by combining a respiratory motion correction approach with temporal regularization in a unique reconstruction algorithm for dynamic PET imaging. Elastic transformation parameters for the motion correction are estimated from the non-attenuation-corrected PET images. The derived displacement matrices are subsequently used in a list-mode based OSEM reconstruction algorithm integrating a temporal regularization between the 3D dynamic PET frames, based on temporal basis functions. These functions are simultaneously estimated at each iteration, along with their relative coefficients for each image voxel. Quantitative evaluation has been performed using dynamic FDG PET/CT acquisitions of lung cancer patients acquired on a GE DRX system. The performance of the proposed method is compared with that of a standard multi-frame OSEM reconstruction algorithm. The proposed method achieved substantial improvements in terms of noise reduction while accounting for loss of contrast due to respiratory motion. Results on simulated data showed that the proposed 4D algorithms led to bias reduction values up to 40% in both tumor and blood regions for similar standard deviation levels, in comparison with a standard 3D reconstruction. Patlak parameter estimations on reconstructed images with the proposed reconstruction methods resulted in 30% and 40% bias reduction in the tumor and lung region respectively for the Patlak slope, and a 30% bias reduction for the intercept in the tumor region (a similar Patlak intercept was achieved in the lung area). Incorporation of the respiratory motion correction using an elastic model along with a temporal regularization in the reconstruction process of the PET dynamic series led to substantial quantitative improvements and motion artifact reduction. Future work will include the integration of a linear FDG kinetic model, in order to directly reconstruct parametric images.

Motion correction using anatomical information in PET/CT and PET/MR hybrid imaging.

Respiratory and cardiac motion causes qualitative and quantitative inaccuracies in whole body multi-modality imaging such as positron emission tomography coupled with computed tomography (PET/CT) and positron emission tomography/magnetic resonance imaging (PET/MRI). Solutions presented to date include motion synchronized PET and corresponding anatomical acquisitions (four dimensional [4D] PET/CT, 4D PET/MR), frequently referred to as the gating approach. This method is based on the acquisition of an external surrogate using an external device (pressure belt, optical monitoring system, spirometer etc.), subsequently used to bin PET and CT or MR anatomical data into a number of gates. A first limitation of this method is the low signal to noise ratio (SNR) of the resulting motion synchronized PET frames, given that every reconstructed frame contains only part of the count statistics available throughout a motion average PET acquisition. Another limitation is that the complex motion of internal organs cannot be fully estimated, characterized and modelled using a mono-dimensional motion signal. In order to resolve such issues, many advanced techniques have been proposed which include three consecutive major steps. These are based on firstly acquiring an external or internal motion surrogate, estimating or modelling the internal motion using anatomical information extracted from 4D anatomical images (CT and/or MR) and finally correcting for motion either in the PET raw data space, the image space or incorporate it within the PET image reconstruction which is the most optimal based motion correction method in PET/CT and in PET/MR imaging. Current research efforts are concentrating on combining the last two steps within a joint motion estimation/motion correction approach, the exploitation of MRI specific motion characterization sequences and the combination of both respiratory and cardiac motion corrections. The goal of this review is to present and discuss the different steps of all these motion correction methods in PET/CT and PET/MR imaging for whole body applications.

A novel partial volume effects correction technique integrating deconvolution associated with denoising within an iterative PET image reconstruction.

PURPOSE: Partial volume effect (PVE) plays an important role in both qualitative and quantitative PET image accuracy, especially for small structures. A previously proposed voxelwise PVE correction method applied on PET reconstructed images involves the use of Lucy-Richardson deconvolution incorporating wavelet-based denoising to limit the associated propagation of noise. The aim of this study is to incorporate the deconvolution, coupled with the denoising step, directly inside the iterative reconstruction process to further improve PVE correction. METHODS: The list-mode ordered subset expectation maximization (OSEM) algorithm has been modified accordingly with the application of the Lucy-Richardson deconvolution algorithm to the current estimation of the image, at each reconstruction iteration. Acquisitions of the NEMA NU2-2001 IQ phantom were performed on a GE DRX PET/CT system to study the impact of incorporating the deconvolution inside the reconstruction [with and without the point spread function (PSF) model] in comparison to its application postreconstruction and to standard iterative reconstruction incorporating the PSF model. The impact of the denoising step was also evaluated. Images were semiquantitatively assessed by studying the trade-off between the intensity recovery and the noise level in the background estimated as relative standard deviation. Qualitative assessments of the developed methods were additionally performed on clinical cases. RESULTS: Incorporating the deconvolution without denoising within the reconstruction achieved superior intensity recovery in comparison to both standard OSEM reconstruction integrating a PSF model and application of the deconvolution algorithm in a postreconstruction process. The addition of the denoising step permitted to limit the SNR degradation while preserving the intensity recovery. CONCLUSIONS: This study demonstrates the feasibility of incorporating the Lucy-Richardson deconvolution associated with a wavelet-based denoising in the reconstruction process to better correct for PVE. Future work includes further evaluations of the proposed method on clinical datasets and the use of improved PSF models.