Normalized Mean Square Error Research Articles

Linear programming-aided power allocation for large OCDMA networks

To provide a suitable operation in optical code division multiple access (OCDMA) networks, it is paramount to balance the powers received at the destination optical node. This work presents a solution strategy for the power allocation (PA) problem in OCDMA by rewriting it as a linear programming (LP) problem and applying two LP methods based on the Simplex method and the Interior Point method (IPM). Such LP methods proposed in the PA OCDMA context were compared with two methods available in the literature: a) hybrid ALPSO PA method, which is based on the particle swarm optimization (PSO) strategy combined with the augmented Lagrangian (AL) analytical method and the solver GUROBI; b) the high-complexity benchmark solution matrix inversion method (MIM), which is used to verify the quality of the Simplex and IPM solutions. Numerical result reveal the effectiveness and efficiency of both LP methods when compared with other competitive methods. Numerical results in terms of floating-point operations (FLOPS), normalized mean squared error (NMSE), convergence, and the evolution of the allocated power reveal the effectiveness and efficiency of both LP methods when compared with the literature methods, mainly under higher network dimensions (K≥32 optical nodes), achieving better accuracy-complexity tradeoffs. For 32≤K≤512 users, the IPM has resulted in perfect feasibility (F=0) and little complexity, i.e., an order less in terms of Flops than Simplex, twice to five times less than MIM procedure, and at least four orders less complexity than the hybrid analytical-heuristic ALPSO method.

Machine Learning-Based Hearing Aid Fitting Personalization Using Clinical Fitting Data.

The initial software fitting prescribed by the fitting formula largely depends on the patient's hearing loss, which may not be the optimal preference for a particular user. Certain criteria must also be readjusted by an audiologist to meet the user-specific requirements. Therefore, this study focuses on the novel application of a neural network (NN) technique to build a suitable fitting algorithm with prescribed hearing loss and the corresponding preferred gain to minimize the gap between optimized fittings. The algorithm intended to learn the hearing preferences of an individual user such that the initial fitting may be optimized. These findings demonstrate the efficiency of the algorithm, with and without additional features. Using the clinical fitting data, the average mean square error (MSE) for the simple NN algorithm was 5.4183%. By adding additional features to the data, the algorithm performed better, and the average MSE was as low as 5.2530%. However, the algorithm outperformed Company A fitting software, as the MSE was the highest at 5.4748%. As the company's automatic fitting has a noticeable discrepancy with clinical fitting records, the impeccable results from this study can lead to a better path towards fitting satisfaction, thus benefiting the hearing-impaired community to a larger extent.

A Comparison of Surrogate Behavioral Models for Power Amplifier Linearization under High Sparse Data.

A good approximation to power amplifier (PA) behavioral modeling requires precise baseband models to mitigate nonlinearities. Since digital predistortion (DPD) is used to provide the PA linearization, a framework is necessary to validate the modeling figures of merit support under signal conditioning and transmission restrictions. A field-programmable gate array (FPGA)-based testbed is developed to measure the wide-band PA behavior using a single-carrier 64-quadrature amplitude modulation (QAM) multiplexed by orthogonal frequency-division multiplexing (OFDM) based on long-term evolution (LTE) as a stimulus, with different bandwidths signals. In the search to provide a heuristic target approach modeling, this paper introduces a feature extraction concept to find an appropriate complexity solution considering the high sparse data issue in amplitude to amplitude (AM-AM) and amplitude to phase AM-PM models extraction, whose penalties are associated with overfitting and hardware complexity in resulting functions. Thus, experimental results highlight the model performance for a high sparse data regime and are compared with a regression tree (RT), random forest (RF), and cubic-spline (CS) model accuracy capabilities for the signal conditioning to show a reliable validation, low-complexity, according to the peak-to-average power ratio (PAPR), complementary cumulative distribution function (CCDF), coefficients extraction, normalized mean square error (NMSE), and execution time figures of merit. The presented models provide a comparison with original data that aid to compare the dimension and robustness for each surrogate model where (i) machine learning (ML)-based and (ii) CS interpolate-based where high sparse data are present, NMSE between the CS interpolated based are also compared to demonstrate the efficacy in the prediction methods with lower convergence times and complexities.

CT-free attenuation correction for dedicated cardiac SPECT using a 3D dual squeeze-and-excitation residual dense network

BackgroundAttenuation correction (AC) using CT transmission scanning enables the accurate quantitative analysis of dedicated cardiac SPECT. However, AC is challenging for SPECT-only scanners. We developed a deep learning-based approach to generate synthetic AC images from SPECT images without AC. MethodsCT-free AC was implemented using our customized Dual Squeeze-and-Excitation Residual Dense Network (DuRDN). 172 anonymized clinical hybrid SPECT/CT stress/rest myocardial perfusion studies were used in training, validation, and testing. Additional body mass index (BMI), gender, and scatter-window information were encoded as channel-wise input to further improve the network performance. ResultsQuantitative and qualitative analysis based on image voxels and 17-segment polar map showed the potential of our approach to generate consistent SPECT AC images. Our customized DuRDN showed superior performance to conventional network design such as U-Net. The averaged voxel-wise normalized mean square error (NMSE) between the predicted AC images by DuRDN and the ground-truth AC images was 2.01 ± 1.01%, as compared to 2.23 ± 1.20% by U-Net. ConclusionsOur customized DuRDN facilitates dedicated cardiac SPECT AC without CT scanning. DuRDN can efficiently incorporate additional patient information and may achieve better performance compared to conventional U-Net.

Learning How to Transfer From Uplink to Downlink via Hyper-Recurrent Neural Network for FDD Massive MIMO

In order to unlock the full advantages of massive multiple-input multiple-output (MIMO) in the downlink, the base station (BS) must leverage information about the downlink fading channels. However, in frequency division duplex (FDD) systems, full channel reciprocity does not hold, and acquiring information about the downlink channels generally requires downlink pilot transmission followed by uplink feedback. Prior work proposed to design pilot transmission, feedback, and channel state information (CSI) estimation, or directly downlink beamforming, via deep learning in an end-to-end manner. While previous work only used downlink pilots in a single slot, in this work, we introduce an enhanced end-to-end design that leverages partial uplink-downlink reciprocity and temporal correlation of the fading processes by utilizing jointly downlink and uplink pilots across multiple time slots. The proposed method is based on a novel deep learning architecture – HyperRNN – that combines hypernetworks and recurrent neural networks (RNNs) to optimize the transfer of long-term invariant channel features from uplink to downlink. Simulation results demonstrate that the HyperRNN achieves a lower normalized mean square error (NMSE) performance in terms of channel estimation, and that it attains a larger achievable sum-rate when applied to multi-user beamforming, as compared to the state of the art.

Generalized superimposed training for RIS-aided cell-free massive MIMO-OFDM networks

In this paper, a generalized superimposed training (GST) is proposed for an uplink cell-free multiple-input multiple-output orthogonal frequency-division multiplexing (mMIMO-OFDM) system, which is aided by reconfigurable intelligent surfaces (RISs) to enhance the spectral efficiency in the system. For the GST scheme, the pilots and data symbols are transmitted simultaneously in the coherence time. This scheme is different from traditional separate transmission methods, such as regular pilots (RP) transmission. In the OFDM multi-carrier case, a part of the subcarriers is based on the GST, whereas the other part of subcarriers is used for data transmission only. The channel and data estimations are carried out and the normalized mean-squared error (NMSE), bit error rate (BER), and sum-rate in different schemes are compared. Different receiver cooperation levels are analyzed in this case, including fully centralized processing and local processing. The distributed time processing and iterative process are also used to improve the performance of the data estimation in this system.

Mobility and Routing based Channel Estimation for Hybrid Millimeter-Wave MIMO WSNs

Recently, technological developments have enhanced, the use of Millimeter-wave (mm-Wave) Multiple Input Multiple Output (MIMO) system in various communication applications and wireless sensor networks as channel estimation efficiency can be immensely improved with the help of this technological developments in Millimeter-wave MIMO system and wireless sensor network as well. Moreover, they can improve quality of communication services to a great extent. However, cell interference in Millimeter-wave (mm-Wave) MIMO system can produce a massive impact on spectral efficiency. Therefore, a Routing Enabled Channel Estimation (RECE) technique is presented in this article to minimize interference between cells. The proposed Channel Estimation technique improves channel capacity as well as spectral efficiency. Moreover, Normalized Mean Square Error (NMSE) is minimized heavily using proposed RECE technique. Here, main aim of this article is to reduce cell interference and channel estimation inside a cell by using route selection, beam selection, and spatial frequency estimation. Here, different scenarios and parameters are considered to evaluate performance efficiency of proposed RECE technique in terms of spectral efficiency, NMSE and SNR and compared against varied traditional channel estimation techniques. Moreover, it is clearly evident from performance results that the proposed channel estimation technique performs better than the other two methods.

Leveraging Machine Learning for Fault-Tolerant Air Pollutants Monitoring for a Smart City Design

Air pollution has become a global issue due to its widespread impact on the environment, economy, civilization and human health. Owing to this, a lot of research and studies have been done to tackle this issue. However, most of the existing methodologies have several issues such as high cost, low deployment, maintenance capabilities and uni-or bi-variate concentration of air pollutants. In this paper, a hybrid CNN-LSTM model is presented to forecast multivariate air pollutant concentration for the Internet of Things (IoT) enabled smart city design. The amalgamation of CNN-LSTM acts as an encoder-decoder which improves the overall accuracy and precision. The performance of the proposed CNN-LSTM is compared with conventional and hybrid machine learning (ML) models on the basis of Mean Absolute Error (MAE), Mean Absolute Percentage Error (MAPE) and Mean Square Error (MSE). The proposed model outperforms various state-of-the-art ML models by generating an average MAE, MAPE and MSE of 54.80%, 52.78% and 60.02%. Furthermore, the predicted results are cross-validated with the actual concentration of air pollutants and the proposed model achieves a high degree of prediction accuracy to real-time air pollutants concentration. Moreover, a cross-grid cooperative scheme is proposed to tackle the IoT monitoring station malfunction scenario and make the pollutant monitoring more fault resistant and robust. The proposed scheme exploits the correlation between neighbouring monitoring stations and air pollutant concentration. The model generates an average MAPE and MSE of 10.90% and 12.02%, respectively.

DuDoSS: Deep-learning-based dual-domain sinogram synthesis from sparsely sampled projections of cardiac SPECT.

Myocardial perfusion imaging (MPI) using single-photon emission-computed tomography (SPECT) is widely applied for the diagnosis of cardiovascular diseases. In clinical practice, the long scanning procedures and acquisition time might induce patient anxiety and discomfort, motion artifacts, and misalignments between SPECT and computed tomography (CT). Reducing the number of projection angles provides a solution that results in a shorter scanning time. However, fewer projection angles might cause lower reconstruction accuracy, higher noise level, and reconstruction artifacts due to reduced angular sampling. We developed a deep-learning-based approach for high-quality SPECT image reconstruction using sparsely sampled projections. We proposed a novel deep-learning-based dual-domain sinogram synthesis (DuDoSS) method to recover full-view projections from sparsely sampled projections of cardiac SPECT. DuDoSS utilized the SPECT images predicted in the image domain as guidance to generate synthetic full-view projections in the sinogram domain. The synthetic projections were then reconstructed into non-attenuation-corrected and attenuation-corrected (AC) SPECT images for voxel-wise and segment-wise quantitative evaluations in terms of normalized mean square error (NMSE) and absolute percent error (APE). Previous deep-learning-based approaches, including direct sinogram generation (Direct Sino2Sino) and direct image prediction (Direct Img2Img), were tested in this study for comparison. The dataset used in this study included a total of 500 anonymized clinical stress-state MPI studies acquired on a GE NM/CT 850 scanner with 60 projection angles following the injection of 99m Tc-tetrofosmin. Our proposed DuDoSS generated more consistent synthetic projections and SPECT images with the ground truth than other approaches. The average voxel-wise NMSE between the synthetic projections by DuDoSS and the ground-truth full-view projections was 2.08%±0.81%, as compared to 2.21%±0.86% (p<0.001) by Direct Sino2Sino. The averaged voxel-wise NMSE between the AC SPECT images by DuDoSS and the ground-truth AC SPECT images was 1.63%±0.72%, as compared to 1.84%±0.79% (p<0.001) by Direct Sino2Sino and 1.90%±0.66% (p<0.001) by Direct Img2Img. The averaged segment-wise APE between the AC SPECT images by DuDoSS and the ground-truth AC SPECT images was 3.87%±3.23%, as compared to 3.95%±3.21% (p=0.023) by Direct Img2Img and 4.46%±3.58% (p<0.001) by Direct Sino2Sino. Our proposed DuDoSS is feasible to generate accurate synthetic full-view projections from sparsely sampled projections for cardiac SPECT. The synthetic projections and reconstructed SPECT images generated from DuDoSS are more consistent with the ground-truth full-view projections and SPECT images than other approaches. DuDoSS can potentially enable fast data acquisition of cardiac SPECT.

Selective Algorithm for Expanded Group Method of Data Handling Applied to Power Amplifier Modeling

Power amplifiers (PAs) are electronic devices commonly used in telecommunications that need to transmit information with high energetic efficiency. For this, it is necessary to use data manipulation methods that assist in the linearization of the output signal. Our previous conference paper presented two codes constructed based on the Group Method of Data Handling (GMDH) and which differ in their way of selecting the best coefficients to be used in the calculations of the neural network. The first method, called Embracing, assumes greater availability of data, while the second, called Selective, selects information from the beginning of the code. This work extends the previous GMDH models by expanding the PA inputs into Laguerre basis functions with a single real pole. The comparison among the different approaches employs experimental data collected from a GaN HEMT class AB PA and a Si LDMOS class AB. The most selective and computationally more complex structure, when searching for identification since from the first layers, expresses minor errors and the best results in the output for both Conventional and Expanded GMDH models, becoming a reasoned option for use in PAs. A normalized mean square error (NMSE) of -35.44 dB was obtained by Expanded GMDH with Selective algorithm and 5 inputs when using the GaN PA, whereas a NMSE of -40.35 dB was obtained by the Expanded GMDH with Selective algorithm and 4 inputs when calculated with the Si LDMOS PA data.

Three-Phase Load Prediction-Based Hybrid Convolution Neural Network Combined Bidirectional Long Short-Term Memory in Solar Power Plant

The economic renewable energy generations have been rapidly developed because of the sharp reduction in the costs of solar panels. It is imperative to forecast the three-phase load power for more effective energy planning and optimization in a smart solar microgrid installed on a building in the Linyuan District, Taiwan. To alleviate this problem, this article proposes a convolution neural network bidirectional long short-term memory (CNN-Bi-LSTM) to accurately predict the short-term three-phase load power in building the energy management system in the smart solar microgrid with the collected data from advanced metering infrastructure (AMI), which have not been investigated before. The three-phase load-predicting methodology is developed using weather parameters and different collected data from AMI. The project evaluates the performance of the CNN-Bi-LSTM model by utilizing hyper-parameter optimization to attain the optimum parameters. The prediction models are trained based on hourly historical input features, selected based on the Pearson correlation coefficient. The performances’ optimal structure CNN-Bi-LSTM are validated and compared with the bidirectional LSTM (Bi-LSTM), LSTM, the Gated Recurrent Unit (GRU), and the recurrent neural network (RNN) models. The obtained optimized structure of CNN-Bi-LSTM demonstrates the effectiveness of the proposed models in the short-term prediction of three-phase load power in a smart solar microgrid for building with a maximum enhancement of 68.36% and 8.81% average MSE, and 30.26% and 36.36% average MAE during the testing and validating operations.

A densely interconnected network for deep learning accelerated MRI

ObjectiveTo improve accelerated MRI reconstruction through a densely connected cascading deep learning reconstruction framework.Materials and methodsA cascading deep learning reconstruction framework (reference model) was modified by applying three architectural modifications: input-level dense connections between cascade inputs and outputs, an improved deep learning sub-network, and long-range skip-connections between subsequent deep learning networks. An ablation study was performed, where five model configurations were trained on the NYU fastMRI neuro dataset with an end-to-end scheme conjunct on four- and eightfold acceleration. The trained models were evaluated by comparing their respective structural similarity index measure (SSIM), normalized mean square error (NMSE), and peak signal to noise ratio (PSNR).ResultsThe proposed densely interconnected residual cascading network (DIRCN), utilizing all three suggested modifications achieved a SSIM improvement of 8% and 11%, a NMSE improvement of 14% and 23%, and a PSNR improvement of 2% and 3% for four- and eightfold acceleration, respectively. In an ablation study, the individual architectural modifications all contributed to this improvement for both acceleration factors, by improving the SSIM, NMSE, and PSNR with approximately 2–4%, 4–9%, and 0.5–1%, respectively.ConclusionThe proposed architectural modifications allow for simple adjustments on an already existing cascading framework to further improve the resulting reconstructions.

Fault Restoration of Six-Axis Force/Torque Sensor Based on Optimized Back Propagation Networks.

Six-axis force/torque sensors are widely installed in manipulators to help researchers achieve closed-loop control. When manipulators work in comic space and deep sea, the adverse ambient environment will cause various degrees of damage to F/T sensors. If the disability of one or two dimensions is restored by self-restoration methods, the robustness and practicality of F/T sensors can be considerably enhanced. The coupling effect is an important characteristic of multi-axis F/T sensors, which implies that all dimensions of F/T sensors will influence each other. We can use this phenomenon to speculate the broken dimension by other regular dimensions. Back propagation neural network (BPNN) is a classical feedforward neural network, which consists of several layers and adopts the back-propagation algorithm to train networks. Hyperparameters of BPNN cannot be updated by training, but they impact the network performance directly. Hence, the particle swarm optimization (PSO) algorithm is adopted to tune the hyperparameters of BPNN. In this work, each dimension of a six-axis F/T sensor is regarded as an element in the input vector, and the relationships among six dimensions can be obtained using optimized BPNN. The average MSE of restoring one dimension and two dimensions over the testing data is and , respectively. Furthermore, the average quote error of one restored dimension and two restored dimensions are and , respectively. The analysis of experimental results illustrates that the proposed fault restoration method based on PSO-BPNN is viable and practical. The F/T sensor restored using the proposed method can reach the original measurement precision.

Comparative Study of the Error Trend and Seasonal Exponential Smoothing and ARIMA Model Using COVID-19 Death Rate in Nigeria

Background: COVID-19 has claimed the lives of millions of people in Nigeria and around the world during the last two years. It is a recognized global health crisis of our day, as well as a persistent threat to the earth. The goal of this study was to examine the trend and fit an Error Trend and Seasonal (ETS) exponential smoothing and Autoregressive Integrated Moving Average (ARIMA) model to Nigeria's COVID-19 daily fatalities.Methods: A dataset of daily COVID-19 confirmed fatality cases was used in the investigation. Data was acquired from the Nigerian Centre for Disease Control (NCDC) web database between the 10th of July 2020 and the 2nd of December 2021. The ARIMA model and twelve (12) ETS exponential smoothing techniques were investigated using a dataset of COVID-19 pandemic deaths in Nigeria. The ARIMA and ETS exponential smoothing algorithms were evaluated using the Akaike Information Criterion (AIC), Bayesian Information Criterion (BIC), Hannan Quinn Information Criterion (HQC), and Average Mean Squared Error (AMSE) selection criteria.Result: The ARIMA (0,1,0) model was the best time series modeling for the coronavirus (COVID-19) epidemic in Nigeria since it had the lowest AIC=2863.51, BIC=2866.90, HQ = 2866.90, and AMSE = 0.55471 values.Conclusion: The ARIMA (0,1,0) model is preferred above the other thirteen (13) competing models based on daily confirmed COVID-19 deaths in Nigeria. This research would assist the Nigerian government in better understanding the pestilence's evolution pattern and providing adequate provisions, prompt mediation, and treatment to prevent additional deaths caused by the virus.

Estimation and Compensation of IQ Imbalance for OTSM Systems

Recently, orthogonal time sequency multiplexing (OTSM) waveform was proposed using Walsh Hadamard transform (WHT) for doubly-dispersive channels at a lower modulation/demodulation complexity than orthogonal time frequency space (OTFS) waveform. At high frequencies, hardware imperfections such as in-phase and quadrature (IQ) imbalance is prominent at the transceiver. In this letter, the input-output (I/O) relation in the delay-sequency domain is derived under these impairments at the receiver. The impairment parameters are estimated in the delay-time domain and are compensated in the time domain using proposed dual pilot approach. The normalized mean square error (NMSE) of the estimated imbalance parameters is evaluated and it is observed that NMSE reduces with increase in SNR and with reduction in mobility. The results are evaluated with OTFS system and it is observed that the NMSE performance of OTSM is similar to OTFS under high pilot power and high SNR at no excess pilot power. The bit error rate performance of the compensated system is evaluated and its performance is close to the ideal system.

A Novel Reservoir Computing Platform Constructed by Laser Controlled GaAs Resonator

In this work, a novel realization of the well known machine learning approach—reservoir computing (RC) using nanoelectromechanical system (NEMS) GaAs resonator is proposed. The specially designed resonator can interact with laser directly without cavity. Thanks to the richness of nonlinearity embedded in the light-mediated dynamical interacting process, the RC platform performs outstandingly even by eliminating the time delay feedback, and the conduction of nonlinear autoregressive moving-average (NARMA) task shows that the normalized mean square error (NMSE) can achieve <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"> <tex-math notation="LaTeX">$2.1\,\,\times10$ </tex-math></inline-formula> <sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">−3</sup> that is one order of magnitude lower than the recently reported result. Detailed theoretical modeling as well as numerical simulations on device’s nonlinear dynamics and RC’s hyperparameters are given. The interplay between the transient dynamics of the resonator and RC performance is particularly investigated.

A Residual-Fitting Modeling Method for Digital Predistortion of Broadband Power Amplifiers

This letter proposes a residual-fitting modeling method for digital predistortion (DPD) of broadband power amplifiers (PAs), and then constructs a residual-fitting model. To avoid directly modeling strong nonlinearity and memory effect, the model is split into a conversion, fitting, and recovery module. In this way, the nonlinearity and memory effect of the output signal of PAs are reduced after the conversion module, and then the fitting module models the converted signal, finally the behavioral characteristics of PAs are recovered by the recovery module. In the experimental test, a 100 MHz orthogonal frequency division multiplexing (OFDM) signal is used as input signal of a Doherty PA. The experimental results show that compared with the existing augmented real-valued time-delay neural network (ARVTDNN), the proposed residual-fitting memory polynomial-ARVTDNN (MP-ARVTDNN) model with much fewer coefficients lowers normalized mean square error (NMSE) and adjacent channel power ratio (ACPR).

TMSCNet: A three-stage multi-branch self-correcting trait estimation network for RGB and depth images of lettuce

Growth traits, such as fresh weight, diameter, and leaf area, are pivotal indicators of growth status and the basis for the quality evaluation of lettuce. The time-consuming, laborious and inefficient method of manually measuring the traits of lettuce is still the mainstream. In this study, a three-stage multi-branch self-correcting trait estimation network (TMSCNet) for RGB and depth images of lettuce was proposed. The TMSCNet consisted of five models, of which two master models were used to preliminarily estimate the fresh weight (FW), dry weight (DW), height (H), diameter (D), and leaf area (LA) of lettuce, and three auxiliary models realized the automatic correction of the preliminary estimation results. To compare the performance, typical convolutional neural networks (CNNs) widely adopted in botany research were used. The results showed that the estimated values of the TMSCNet fitted the measurements well, with coefficient of determination (R2) values of 0.9514, 0.9696, 0.9129, 0.8481, and 0.9495, normalized root mean square error (NRMSE) values of 15.63, 11.80, 11.40, 10.18, and 14.65% and normalized mean squared error (NMSE) value of 0.0826, which was superior to compared methods. Compared with previous studies on the estimation of lettuce traits, the performance of the TMSCNet was still better. The proposed method not only fully considered the correlation between different traits and designed a novel self-correcting structure based on this but also studied more lettuce traits than previous studies. The results indicated that the TMSCNet is an effective method to estimate the lettuce traits and will be extended to the high-throughput situation. Code is available at https://github.com/lxsfight/TMSCNet.git.

Graph neural networks for multivariate time series regression with application to seismic data

Machine learning, with its advances in deep learning has shown great potential in analyzing time series. In many scenarios, however, additional information that can potentially improve the predictions is available. This is crucial for data that arise from e. g., sensor networks that contain information about sensor locations. Then, such spatial information can be exploited by modeling it via graph structures, along with the sequential (time series) information. Recent advances in adapting deep learning to graphs have shown potential in various tasks. However, these methods have not been adapted for time series tasks to a great extent. Most attempts have essentially consolidated around time series forecasting with small sequence lengths. Generally, these architectures are not well suited for regression or classification tasks where the value to be predicted is not strictly depending on the most recent values, but rather on the whole length of the time series. We propose TISER-GCN, a novel graph neural network architecture for processing, in particular, these long time series in a multivariate regression task. Our proposed model is tested on two seismic datasets containing earthquake waveforms, where the goal is to predict maximum intensity measurements of ground shaking at each seismic station. Our findings demonstrate promising results of our approach—with an average MSE reduction of 16.3%—compared to the best performing baselines. In addition, our approach matches the baseline scores by needing only half the input size. The results are discussed in depth with an additional ablation study.

Virtual high-count PET image generation using a deep learning method.

Recently, deep learning-based methods have been established to denoise the low-count positron emission tomography (PET) images and predict their standard-count image counterparts, which could achieve reduction of injected dosage and scan time, and improve image quality for equivalent lesion detectability and clinical diagnosis. In clinical settings, the majority scans are still acquired using standard injection dose with standard scan time. In this work, we applied a 3D U-Net network to reduce the noise of standard-count PET images to obtain the virtual-high-count (VHC) PET images for identifying the potential benefits of the obtained VHC PETimages. The training datasets, including down-sampled standard-count PET images as the network input and high-count images as the desired network output, were derived from 27 whole-body PET datasets, which were acquired using 90-min dynamic scan. The down-sampled standard-count PET images were rebinned with matched noise level of 195 clinical static PET datasets, by matching the normalized standard derivation (NSTD) inside 3D liver region of interests (ROIs). Cross-validation was performed on 27 PET datasets. Normalized mean square error (NMSE), peak signal to noise ratio (PSNR), structural similarity index (SSIM), and standard uptake value (SUV) bias of lesions were used for evaluation on standard-count and VHC PET images, with real-high-count PET image of 90 min as the gold standard. In addition, the network trained with 27 dynamic PET datasets was applied to 195 clinical static datasets to obtain VHC PET images. The NSTD and mean/max SUV of hypermetabolic lesions in standard-count and VHC PET images were evaluated. Three experienced nuclear medicine physicians evaluated the overall image quality of randomly selected 50 out of 195 patients' standard-count and VHC images and conducted 5-score ranking. A Wilcoxon signed-rank test was used to compare differences in the grading of standard-count and VHCimages. The cross-validation results showed that VHC PET images had improved quantitative metrics scores than the standard-count PET images. The mean/max SUVs of 35 lesions in the standard-count and true-high-count PET images did not show significantly statistical difference. Similarly, the mean/max SUVs of VHC and true-high-count PET images did not show significantly statistical difference. For the 195 clinical data, the VHC PET images had a significantly lower NSTD than the standard-count images. The mean/max SUVs of 215 hypermetabolic lesions in the VHC and standard-count images showed no statistically significant difference. In the image quality evaluation by three experienced nuclear medicine physicians, standard-count images and VHC images received scores with mean and standard deviation of 3.34±0.80 and 4.26 ± 0.72 from Physician 1, 3.02 ± 0.87 and 3.96 ± 0.73 from Physician 2, and 3.74 ± 1.10 and 4.58 ± 0.57 from Physician 3, respectively. The VHC images were consistently ranked higher than the standard-count images. The Wilcoxon signed-rank test also indicated that the image quality evaluation between standard-count and VHC images had significantdifference. A DL method was proposed to convert the standard-count images to the VHC images. The VHC images had reduced noise level. No significant difference in mean/max SUV to the standard-count images was observed. VHC images improved image quality for better lesion detectability and clinicaldiagnosis.