Dataset Information

A deep neural network for fast and accurate scatter estimation in quantitative SPECT/CT under challenging scatter conditions.

ABSTRACT: PURPOSE:A major challenge for accurate quantitative SPECT imaging of some radionuclides is the inadequacy of simple energy window-based scatter estimation methods, widely available on clinic systems. A deep learning approach for SPECT/CT scatter estimation is investigated as an alternative to computationally expensive Monte Carlo (MC) methods for challenging SPECT radionuclides, such as ⁹⁰Y. METHODS:A deep convolutional neural network (DCNN) was trained to separately estimate each scatter projection from the measured ⁹⁰Y bremsstrahlung SPECT emission projection and CT attenuation projection that form the network inputs. The 13-layer deep architecture consisted of separate paths for the emission and attenuation projection that are concatenated before the final convolution steps. The training label consisted of MC-generated "true" scatter projections in phantoms (MC is needed only for training) with the mean square difference relative to the model output serving as the loss function. The test data set included a simulated sphere phantom with a lung insert, measurements of a liver phantom, and patients after ⁹⁰Y radioembolization. OS-EM SPECT reconstruction without scatter correction (NO-SC), with the true scatter (TRUE-SC) (available for simulated data only), with the DCNN estimated scatter (DCNN-SC), and with a previously developed MC scatter model (MC-SC) were compared, including with ⁹⁰Y PET when available. RESULTS:The contrast recovery (CR) vs. noise and lung insert residual error vs. noise curves for images reconstructed with DCNN-SC and MC-SC estimates were similar. At the same noise level of 10% (across multiple realizations), the average sphere CR was 24%, 52%, 55%, and 67% for NO-SC, MC-SC, DCNN-SC, and TRUE-SC, respectively. For the liver phantom, the average CR for liver inserts were 32%, 73%, and 65% for NO-SC, MC-SC, and DCNN-SC, respectively while the corresponding values for average contrast-to-noise ratio (visibility index) in low-concentration extra-hepatic inserts were 2, 19, and 61, respectively. In patients, there was high concordance between lesion-to-liver uptake ratios for SPECT reconstruction with DCNN-SC (median 4.8, range 0.02-13.8) compared with MC-SC (median 4.0, range 0.13-12.1; CCC?=?0.98) and with ⁹⁰Y PET (median 4.9, range 0.02-11.2; CCC?=?0.96) while the concordance with NO-SC was poor (median 2.8, range 0.3-7.2; CCC?=?0.59). The trained DCNN took ~?40 s (using a single i5 processor on a desktop computer) to generate the scatter estimates for all 128 views in a patient scan, compared to ~?80 min for the MC scatter model using 12 processors. CONCLUSIONS:For diverse ⁹⁰Y test data that included patient studies, we demonstrated comparable performance between images reconstructed with deep learning and MC-based scatter estimates using metrics relevant for dosimetry and for safety. This approach that can be generalized to other radionuclides by changing the training data is well suited for real-time clinical use because of the high speed, orders of magnitude faster than MC, while maintaining high accuracy.

SUBMITTER: Xiang H

PROVIDER: S-EPMC7666660 | biostudies-literature | 2020 Dec

REPOSITORIES: biostudies-literature

ACCESS DATA

Publications

A deep neural network for fast and accurate scatter estimation in quantitative SPECT/CT under challenging scatter conditions.

Xiang Haowei H Lim Hongki H Fessler Jeffrey A JA Dewaraja Yuni K YK

European journal of nuclear medicine and molecular imaging 20200515 13

<h4>Purpose</h4>A major challenge for accurate quantitative SPECT imaging of some radionuclides is the inadequacy of simple energy window-based scatter estimation methods, widely available on clinic systems. A deep learning approach for SPECT/CT scatter estimation is investigated as an alternative to computationally expensive Monte Carlo (MC) methods for challenging SPECT radionuclides, such as <sup>90</sup>Y.<h4>Methods</h4>A deep convolutional neural network (DCNN) was trained to separately es ...[more]

PMID: 32415551

Similar Datasets

Project description:PURPOSE:In 90 Y microsphere radioembolization (RE), accurate post-therapy imaging-based dosimetry is important for establishing absorbed dose versus outcome relationships for developing future treatment planning strategies. Additionally, accurately assessing microsphere distributions is important because of concerns for unexpected activity deposition outside the liver. Quantitative 90 Y imaging by either SPECT or PET is challenging. In 90 Y SPECT model based methods are necessary for scatter correction because energy window-based methods are not feasible with the continuous bremsstrahlung energy spectrum. The objective of this work was to implement and evaluate a scatter estimation method for accurate 90 Y bremsstrahlung SPECT/CT imaging. METHODS:Since a fully Monte Carlo (MC) approach to 90 Y SPECT reconstruction is computationally very demanding, in the present study the scatter estimate generated by a MC simulator was combined with an analytical projector in the 3D OS-EM reconstruction model. A single window (105 to 195-keV) was used for both the acquisition and the projector modeling. A liver/lung torso phantom with intrahepatic lesions and low-uptake extrahepatic objects was imaged to evaluate SPECT/CT reconstruction without and with scatter correction. Clinical application was demonstrated by applying the reconstruction approach to five patients treated with RE to determine lesion and normal liver activity concentrations using a (liver) relative calibration. RESULTS:There was convergence of the scatter estimate after just two updates, greatly reducing computational requirements. In the phantom study, compared with reconstruction without scatter correction, with MC scatter modeling there was substantial improvement in activity recovery in intrahepatic lesions (from > 55% to > 86%), normal liver (from 113% to 104%), and lungs (from 227% to 104%) with only a small degradation in noise (13% vs. 17%). Similarly, with scatter modeling contrast improved substantially both visually and in terms of a detectability index, which was especially relevant for the low uptake extrahepatic objects. The trends observed for the phantom were also seen in the patient studies where lesion activity concentrations and lesion-to-liver concentration ratios were lower for SPECT without scatter correction compared with reconstruction with just two MC scatter updates: in eleven lesions the mean uptake was 4.9 vs. 7.1 MBq/mL (P = 0.0547), the mean normal liver uptake was 1.6 vs. 1.5 MBq/mL (P = 0.056) and the mean lesion-to-liver uptake ratio was 2.7 vs. 4.3 (P = 0.0402) for reconstruction without and with scatter correction respectively. CONCLUSIONS:Quantitative accuracy of 90 Y bremsstrahlung imaging can be substantially improved with MC scatter modeling without significant degradation in image noise or intensive computational requirements.

Project description:BackgroundThe introduction of hybrid SPECT/CT devices enables quantitative imaging in SPECT, providing a methodological setup for quantitation using SPECT tracers comparable to PET/CT. We evaluated a specific quantitative reconstruction algorithm for SPECT data using a 99mTc-filled NEMA phantom. Quantitative and qualitative image parameters were evaluated for different parametrizations of the acquisition and reconstruction protocol to identify an optimized quantitative protocol.ResultsThe reconstructed activity concentration (ACrec) and the signal-to-noise ratio (SNR) of all examined protocols (n = 16) were significantly affected by the parametrization of the weighting factor k used in scatter correction, the total number of iterations and the sphere volume (all, p < 0.0001). The two examined SPECT acquisition protocols (with 60 or 120 projections) had a minor impact on the ACrec and no significant impact on the SNR. In comparison to the known AC, the use of default scatter correction (k = 0.47) or object-specific scatter correction (k = 0.18) resulted in an underestimation of ACrec in the largest sphere volume (26.5 ml) by - 13.9 kBq/ml (- 16.3%) and - 7.1 kBq/ml (- 8.4%), respectively. An increase in total iterations leads to an increase in estimated AC and a decrease in SNR. The mean difference between ACrec and known AC decreased with an increasing number of total iterations (e.g., for 20 iterations (2 iterations/10 subsets) = - 14.6 kBq/ml (- 17.1%), 240 iterations (24i/10s) = - 8.0 kBq/ml (- 9.4%), p < 0.0001). In parallel, the mean SNR decreased significantly from 2i/10s to 24i/10s by 76% (p < 0.0001).ConclusionQuantitative SPECT imaging is feasible with the used reconstruction algorithm and hybrid SPECT/CT, and its consistent implementation in diagnostics may provide perspectives for quantification in routine clinical practice (e.g., assessment of bone metabolism). When combining quantitative analysis and diagnostic imaging, we recommend using two different reconstruction protocols with task-specific optimized setups (quantitative vs. qualitative reconstruction). Furthermore, individual scatter correction significantly improves both quantitative and qualitative results.

Project description:A well-established approach for diagnostic imaging of osteomyelitis (OM), a bone infection, is simultaneous SPECT-CT of 99mTc sulfur colloid (SC) and 111In white blood cells (WBC). This method provides essentially perfect spatial registration of the tracers within anatomic sites of interest. Currently, diagnosis is based purely on a visual assessment-where relative discordance between 99mTc and 111In uptake in bone, i.e., high 111In and low 99mTc, suggests OM. To achieve more quantitative images, noise, scatter, and crosstalk between radionuclides must be addressed through reconstruction. Here the authors compare their Monte Carlo-based joint OSEM (MC-JOSEM) algorithm, which reconstructs both radionuclides simultaneously, to a more conventional triple-energy window-based reconstruction (TEW-OSEM), and to iterative reconstruction with no compensation for scatter (NC-OSEM).The authors created numerical phantoms of the foot and torso. Multiple bone-infection sites were modeled using high-count Monte Carlo simulation. Counts per voxel were then scaled to values appropriate for 111In WBC and 99mTc SC imaging. Ten independent noisy projection image sets were generated by drawing random Poisson deviates from these very low-noise images. Data were reconstructed using the two iterative scatter-compensation methods, TEW-OSEM and MC-JOSEM, as well as the uncorrected method (NC-OSEM). Mean counts in volumes of interest (VOIs) were used to evaluate the bias and precision of each method. Data were also acquired using a phantom, approximately the size of an adult ankle, consisting of regions representing infected and normal bone marrow, within a bone-like attenuator and surrounding soft tissue; each compartment contained a mixture of 111In and 99mTc. Low-noise data were acquired during multiple short scans over 29 h on a Siemens Symbia T6 SPECT-CT with medium-energy collimators. Pure 99mTc and 111In projection datasets were derived by fitting the acquired projections to the sum of 99mTc and 111In contributions, using the known half-lives. Uncontaminated data were scaled and recombined into six datasets with different activity ratios; ten Poisson noise realizations were then generated for each ratio. VOIs in each of the compartments were used to evaluate the bias and precision of each method with respect to reconstructions of uncontaminated datasets. In addition to the simulated and acquired phantom images, the authors reconstructed patient images with MC-JOSEM and TEW-OSEM. Patient reconstructions were assessed qualitatively for lesion contrast, spatial definition, and scatter.For all simulated and acquired infection phantoms, the root-mean squared-error of measured 99mTc activity was significantly improved with MC-JOSEM and TEW-OSEM in comparison to NC-OSEM reconstructions. While MC-JOSEM trended toward outperforming TEW-OSEM, the improvement was only found to be significant (p<0.001) for the acquired bone phantom in which a wide range of 111In∕99mTc concentration ratios were tested. In all cases, scatter correction did not significantly improve 111In quantitation.Compensation for scatter and crosstalk is useful for improving quality, bias, and precision of 99mTc activity estimates in simultaneous dual-radionuclide imaging of OM. The use of the more rigorous MC-based estimates provided marginal improvements over TEW. While the phantom results were encouraging, more subjects are needed to evaluate the usefulness of quantitative 111In∕99mTc SPECT-CT in the clinic.

Project description:BackgroundFour state-of-the-art single-photon emission computed tomography-computed tomography (SPECT-CT) systems, namely Philips Brightview, General Electric Discovery NM/CT 670 and Infinia Hawkeye 4, and Siemens Symbia T6, were investigated in terms of accuracy of attenuation and scatter correction, contrast recovery for small hot and cold structures, and quantitative capabilities when using their dedicated three-dimensional iterative reconstruction with attenuation and scatter corrections and resolution recovery.MethodsThe National Electrical Manufacturers Association (NEMA) NU-2 1994 phantom with cold air, water, and Teflon inserts, and a homemade contrast phantom with hot and cold rods were filled with 99mTc and scanned. The acquisition parameters were chosen to provide adequate linear and angular sampling and high count statistics. The data were reconstructed using Philips Astonish, General Electric Evolution for Bone, or Siemens Flash3D, eight subsets, and a varying number of iterations. A procedure similar to the one used in positron emission tomography (PET) allowed us to obtain the factor to convert counts per pixel into activity per unit volume.ResultsEdge and oscillation artifacts were observed with all phantoms and all systems. At 30 iterations, the residual fraction in the inserts of the NEMA phantom fell below 3.5%. Contrast recovery increased with the number of iterations but became almost saturated at 24 iterations onwards. In the uniform part of the NEMA and contrast phantoms, a quantification error below 10% was achieved.ConclusionsIn objects whose dimensions exceeded the SPECT spatial resolution by several times, quantification seemed to be feasible within 10% error limits. A partial volume effect correction strategy remains necessary for the smallest structures. The reconstruction artifacts nevertheless remain a handicap on the road towards accurate quantification in SPECT and should be the focus of further works in reconstruction tomography.

Project description:Virtual imaging trials (VITs), defined as the process of conducting clinical imaging trials using computer simulations, offer a time- and cost-effective alternative to traditional imaging trials for CT. The clinical potential of VITs hinges on the realism of simulations modeling the image acquisition process, where the accurate scanner-specific simulation of scatter in a time-feasible manner poses a particular challenge. To meet this need, this study proposes, develops, and validates a rapid scatter estimation framework, based on GPU-accelerated Monte Carlo (MC) simulations and denoising methods, for estimating scatter in single source, dual-source, and photon-counting CT. A CT simulator was developed to incorporate parametric models for an anti-scatter grid and a curved energy integrating detector with an energy-dependent response. The scatter estimates from the simulator were validated using physical measurements acquired on a clinical CT system using the standard single-blocker method. The MC simulator was further extended to incorporate a pre-validated model for a PCD and an additional source-detector pair to model cross scatter in dual-source configurations. To estimate scatter with desirable levels of statistical noise using a manageable computational load, two denoising methods using a (1) convolutional neural network and an (2) optimized Gaussian filter were further deployed. The viability of this framework for clinical VITs was assessed by integrating it with a scanner-specific ray-tracer program to simulate images for an image quality (Mercury) and an anthropomorphic phantom (XCAT). The simulated scatter-to-primary ratios agreed with physical measurements within 4.4% ± 10.8% across all projection angles and kVs. The differences of ∼121 HU between images with and without scatter, signifying the importance of scatter for simulating clinical images. The denoising methods preserved the magnitudes and trends observed in the reference scatter distributions, with an averaged rRMSE value of 0.91 and 0.97 for the two methods, respectively. The execution time of ∼30 s for simulating scatter in a single projection with a desirable level of statistical noise indicates a major improvement in performance, making our tool an eligible candidate for conducting extensive VITs spanning multiple patients and scan protocols.

Dataset Information

A deep neural network for fast and accurate scatter estimation in quantitative SPECT/CT under challenging scatter conditions.

Publications

A deep neural network for fast and accurate scatter estimation in quantitative SPECT/CT under challenging scatter conditions.

Similar Datasets

OmicsDI is part of the ELIXIR infrastructure

Tweets