Kenya Murase

Purpose: To quantitatively evaluate the performance of two types of recurrent neural networks (RNNs), long short‐term memory (LSTM) and gated recurrent units (GRU), using Monte Carlo dropout (MCD) to predict pharmacokinetic (PK) parameters from dynamic contrast‐enhanced magnetic resonance imaging (DCE‐MRI) data.Methods: DCE‐MRI data for simulation studies were synthesized using the extended Tofts model and a population‐averaged arterial input function (AIF). The ranges of PK parameters for training the RNNs were determined from data of patients with brain tumors. The effects of the number of training samples, number of hidden units, dropout rate (DR), and bolus arrival time delay and dispersion in AIF on the accuracy of the PK parameters were investigated, and the uncertainties for different DRs and peak signal‐to‐noise ratios (PSNRs) were quantified. For comparison, PK parameters were estimated using the nonlinear least‐squares method. In the clinical studies, the PK parameter and uncertainty images were generated by applying the trained RNNs to DCE‐MRI data.Results: Compared with GRU, the computational cost for training the LSTM was significantly higher. The prediction accuracy of GRU decreased with decreasing numbers of training samples and hidden units, whereas the performance of LSTM remained stable. Despite an increased computational cost, MCD reduced the prediction error at low PSNR and improved the quality of PK parameter images. The simulation results recommended using a DR of 0.25–0.5 at low PSNR and ≤ 0.25 for other PSNRs. The clinical studies recommended using a DR of 0.25 and 0.5 for LSTM and GRU, respectively.Conclusions: MCD is effective in quantifying uncertainty in PK parameter prediction from DCE‐MRI data and improves their performance, particularly at low PSNR; however, at the expense of increased computational cost. This study helps deepen our understanding of RNNs with MCD and select suitable hyperparameters for creating an RNN architecture for DCE‐MRI studies.

The purpose of this study was to quantitatively evaluate the performance of different simultaneous spatial and temporal regularizers in liver perfusion studies using low-dose dynamic contrast-enhanced computed tomography (DCE-CT). A digital liver phantom was used to simulate chronic liver disease (CLD) and hepatocellular carcinoma (HCC) based on clinical data. Low-dose DCE-CT images were reconstructed using regularizers and a primal-dual algorithm. Subsequently, hepatic perfusion parameter (HPP) images were generated using a dual-input single-compartment model and linear least-squares method. In the CLD model, the effect of regularizers on the input functions (IFs) was examined by calculating the areas under the curves (AUCs) of the IFs, and the HPP estimation accuracy was evaluated by calculating the error and coefficient of variation (CV) between the HPP values obtained by the above methods and true values. In the HCC model, the ratios of the mean HPP values inside and outside the tumor were calculated. The AUCs of IFs decreased with increasing regularization parameter (RP) values. Although the AUC of arterial IF did not significantly depend on the regularizers, that of portal IF did. The error and CV were reduced using low-rank and sparse decomposition (LRSD). Total generalized variation (TGV) combined with LRSD (LTGV) was generally superior to the other regularizers in terms of HPP estimation accuracy and range of available RP values in both the CLD and HCC models. However, striped artifacts were more remarkable in the HPP images obtained by the TGV and LTGV than in those obtained by the other regularizers. The results suggest that the LRSD and LTGV are useful for improving the accuracy of HPP estimation using low-dose DCE-CT and for enhancing its practicality. This study will help select a suitable regularizer and/or RP value for low-dose DCE-CT liver perfusion studies.

Purpose: To apply total generalized variation (TGV) and its combination with low‐rank and sparse decomposition (LRSD) (LTGV) to cerebral perfusion studies using low‐dose dynamic contrast‐enhanced (DCE) CT and to quantitatively evaluate their performances through comparisons with those without any regularizers and those of total variation (TV) and its combination with LRSD (LTV) using simulation and clinical data.Methods: The simulation study used a realistic digital brain phantom. Low‐dose DCE‐CT images were reconstructed using the regularizers and primal‐dual algorithm. Subsequently, cerebral perfusion parameter (CPP) images were generated from them. Thereafter, their quality was evaluated based on the peak signal‐to‐noise ratio (PSNR) and structural similarity index measure (SSIM). Further, the accuracy of CPP estimation was evaluated through a linear regression analysis between the CPP values obtained by the above regularizers and those obtained from the noise‐free DCE‐CT images. In addition, the mean and standard deviation of the CPP were calculated (region analysis). In the clinical study, low‐dose DCE‐CT images were generated using normal‐dose images acquired from a patient, and CPP images were generated from them similar to that in the simulation study.Results: When using LTV and LTGV, both PSNR and SSIM were higher than those of the other methods with increasing regularization parameter values. The results of the linear regression and region analyses demonstrated that TGV generally exhibited the best performance, followed by LTGV, and finally that of TV was significantly different from those of the other regularizers. Despite an overall consistency between the simulation and clinical results, certain inconsistencies appeared owing to the difference in generating low‐dose DCE‐CT images.Conclusions: The results implied that TGV and LTGV were useful in improving the accuracy of CPP estimation using low‐dose DCE‐CT. This study provides an improved understanding of the performance of regularizers and is expected to aid in the selection of a suitable regularizer for low‐dose DCE‐CT perfusion studies.

This study aimed to propose a denoising method for dynamic contrast-enhanced computed tomography (DCE-CT) perfusion studies using a three-dimensional deep image prior (DIP), and to investigate its usefulness in comparison with total variation (TV)-based methods with different regularization parameter (alpha) values through simulation studies. In the proposed DIP method, the DIP was incorporated into the constrained optimization problem for image denoising as a simultaneous spatial and temporal regularizer, which was solved using the alternating direction method of multipliers. In the simulation studies, DCE-CT images were generated using a digital brain phantom and their noise level was varied using the X-ray exposure noise model with different exposures (15, 30, 50, 75, and 100 mAs). Cerebral blood flow (CBF) images were generated from the original contrast enhancement (CE) images and those obtained by the DIP and TV methods using block-circulant singular value decomposition. The quality of the CE images was evaluated using the peak signal-to-noise ratio (PSNR) and structural similarity index (SSIM). To compare the CBF images obtained by the different methods and those generated from the ground truth images, linear regression analysis was performed. When using the DIP method, the PSNR and SSIM were not significantly dependent on the exposure, and the SSIM was the highest for all exposures. When using the TV methods, they were significantly dependent on the exposure and alpha values. The results of the linear regression analysis suggested that the linearity of the CBF images obtained by the DIP method was superior to those obtained from the original CE images and by the TV methods. Our preliminary results suggest that the DIP method is useful for denoising DCE-CT images at ultra-low to low exposures and for improving the accuracy of the CBF images generated from them.

We developed a new image-restoration method that incorporates the point spread function (PSF) into the simultaneous algebraic reconstruction technique (SART-PSF). Additionally, through simulation studies, we investigated the usefulness of the method in comparison with the Richardson–Lucy (RL) algorithm. In the simulation studies, degraded images were generated by convolving magnetic resonance imaging-based brain images with PSF and adding Gaussian or Poisson noise to them to simulate various noise levels. The effects of the number of iterations N, noise, and PSF error on the processed images were quantitatively evaluated using the percent root mean square error (PRMSE) and mean structural similarity index (mSSIM). After applying the SART-PSF to images degraded using Gaussian noise, the PRMSE value and increase thereof, when N was increased, were smaller than those when using the RL algorithm. The mSSIM value was higher and its decrease upon increasing N was smaller than that of the RL algorithm. When Poisson noise was assumed, the differences in PRMSE and mSSIM between both methods were smaller than those when Gaussian noise was assumed. When the PSF error was negative, its effect on PRMSE and mSSIM was similar for both methods. However, when it was positive, the deterioration of these parameters for the SART-PSF was less than that for the RL algorithm in both the Gaussian and Poisson noise cases. The results suggest that the SART-PSF is more robust against noise and a PSF error than the RL algorithm and, thus, can be used as an alternative to the RL algorithm.

The purpose of this study was to develop a method for tissue elasticity imaging using tagged magnetic resonance imaging (MRI). First, we developed a cyclic pressure device that used air to remotely transmit the power to generate cyclic deformation in an object. The pressure induced by the cyclic pressure device was measured by MRI-compatible force sensors. Second, we developed a software to calculate Young's modulus from tagged MRI data using the harmonic phase (HARP) method and the finite element method (FEM). We also developed a software to extract tag-cross points from tagged MRI data. Finally, we evaluated the usefulness of our method using three homogeneous silicone gel phantoms with different degrees of stiffness in comparison with Young's moduli measured by a material testing machine. The coefficient of variation of the pressure data measured by MRI-compatible force sensors was within 5 %, indicating that the reproducibility of the pressure generated by our cyclic pressure device was good. The Young's moduli obtained by the material testing machine, the HARP method, and the FEM increased with increasing stiffness of the phantoms. There were relatively good agreements among the three methods. These results suggest that our method is useful for tissue elasticity imaging and for quantifying the stiffness of tissues.

Purpose: The purpose of this study was to develop a method to simultaneously correct the spatial resolution and inhomogeneous sensitivity of a receiving coil in projection‐based magnetic particle imaging — and to investigate its efficacy through simulation and experimental studies.Methods: Magnetic particle imaging (MPI) images were reconstructed using the simultaneous algebraic reconstruction technique (SART), and simultaneous corrections to sensitivity and spatial resolution were performed by incorporating the sensitivity map of the receiving coil and the system function into the SART algorithm. After each SART update, the regularization method — with total variation (TV) minimization — was used to suppress noise amplification and artifact generation. For comparison, MPI images were also reconstructed using the filtered backprojection (FBP) method and the FBP‐truncated singular value decomposition (TSVD) method, in which the system function was deconvolved from the projection data using TSVD. In simulation studies, the sensitivity map of a second‐order, gradiometer‐type receiving coil was generated using the Biot–Savart law, while the system function was obtained by calculating the MPI signals induced by magnetic nanoparticles at various distances from a field‐free line (FFL), using a lock‐in‐amplifier model. The effects of a regularization parameter for TV minimization (α), number of iterations (N), and signal‐to‐noise ratio (SNR) of the MPI signals on the reconstructed MPI images of a numerical phantom were evaluated, using the image profiles and percent root mean square error (PRMSE). Experimental studies involved the calculation of the system function using a tube phantom. Projection data for an A‐shaped phantom were acquired using our MPI scanner, and their MPI images were reconstructed from the projection data, as described above.Results: When both the sensitivity and spatial resolution were corrected (SART‐SR), the quality of the reconstructed images was seen to have improved, compared to when the spatial resolution was not corrected — or when the FBP and FBP‐TSVD methods were used. When SNR was low (20), a larger α value yielded a better image. The minimum PRMSE occurred at N ≈ 200–400 and increased with increasing N thereafter. When SNR was high (100), the image quality was generally not dependent on the α value within its studied range. The PRMSE decreased slowly with increasing N, and tended to converge to a constant value. The full width at half maximum (FWHM) of the profile was obtained from the A‐shaped phantom, reconstructed using the SART‐SR algorithm with α = 0.05 and N = 1000. The FWHM value of the tube (2 mm diameter) in the A‐shaped phantom image was found to be 2.2 mm on average, whereas those calculated from the images obtained by the FBP and FBP‐TSVD methods were 4.4 and 3.0 mm on average, respectively. Spatial resolution improved when using the FBP‐TSVD method as compared to the FBP method but image distortion and artifacts were observed.Conclusions: Although further studies are necessary to optimize the parameters used in the SART algorithm and in TV minimization, the present results suggest that the proposed method is useful for improving the image quality of projection‐based MPI.

We previously presented a lock-in-amplifier model for analyzing the behavior of signal harmonics in magnetic particle imaging (MPI). In that study, the magnetization and particle size distribution of magnetic nanoparticles (MNPs) were assumed to obey the Langevin theory of paramagnetism and a log-normal distribution, respectively, and a drive magnetic field was assumed to be applied parallel to a selection magnetic field. The purpose of the current study was to investigate the behavior of signal harmonics in MPI with the drive magnetic field being applied to the selection magnetic field in an arbitrary direction using the lock-in-amplifier model. In the lock-in-amplifier model, the signal induced by MNPs in a receiving coil was multiplied with a reference signal, and was then fed through a low-pass filter to extract the DC component of the filtered signal (MPI signal). The strength of the selection magnetic field in MPI was assumed to be given by the product of the gradient strength of the selection magnetic field and the distance from the field-free region (x). The relationships between the MPI signal and x were calculated for odd- and even-numbered harmonics and were investigated for various angles between the axis of a receiving coil and the selection magnetic field. The behavior of signal harmonics in MPI largely depended on the angle between them. This study will be useful for improved understanding, optimization, and development of MPI.

We previously developed a system for projection-based magnetic particle imaging (MPI) with a field-free-line (FFL) encoding scheme. In the projection-based MPI, projection data are given by the convolution between the system function in the spatial domain and the line integral of the concentration of magnetic nanoparticles (MNPs) through the FFL. Thus, it is important to estimate the system functions and to investigate the factors affecting them for enhancing the quantitative property of the projection-based MPI. The purpose of this study was to present a method for generating the system function maps in projection-based MPI. In the simulation studies, the MPI signals induced by MNPs in a receiving coil were calculated using a lock-in-amplifier model under the assumption that the magnetization and particle size distribution of MNPs obey the Langevin theory of paramagnetism and a log-normal distribution, respectively. The system function maps were generated by calculating the MPI signals at various distances from the FFL and angles between the axis of the receiving coil and the selection magnetic field. The effects of the particle size of MNPs, the viscosity of the suspending medium, the amplitude of the drive magnetic field, and the gradient strength of the selection magnetic field on the system function were investigated. The spatial distributions of the system functions could be well understood from their maps generated by our method and were largely affected by the parameters described above. Our method will be useful for improved understanding, optimization, and development of projection-based MPI.