Direct Image-Based Attenuation Correction using Conditional Generative Adversarial Network for SPECT Myocardial Perfusion Imaging.

<span lang="EN-US">Electrical impedance tomography (EIT), as a non-ionizing tomography method, has been widely used in various fields of application, such as engineering and medical fields. This study applies an iterative process to reconstruct EIT images using the simultaneous algebraic reconstruction technique (SART) algorithm combined with K-means clustering. The reconstruction started with defining the finite element method (FEM) model and filtering the measurement data with a Butterworth low-pass filter. The next step is solving the inverse problem in the EIT case with the SART algorithm. The results of the SART algorithm approach were classified using the K-means clustering and thresholding. The reconstruction results were evaluated with the peak signal noise ratio (PSNR), structural similarity indices (SSIM), and normalized root mean square error (NRMSE). They were compared with the one-step gauss-newton (GN) and total variation regularization based on iteratively reweighted least-squares (TV-IRLS) methods. The evaluation shows that the average PSNR and SSIM of the proposed reconstruction method are the highest of the other methods, each being 24.24 and 0.94; meanwhile, the average NRMSE value is the lowest, which is 0.04. The performance evaluation also shows that the proposed method is faster than the other methods.</span>

Objective. To improve positron emission tomography (PET) image quality, we aim to generate images of quality comparable to standard scan duration images using short scan duration (1/8 and 1/16 standard scan duration) inputs and assess the generated standard scan duration images quantitative and qualitatively. Also, the effect of training dataset properties (i.e. body mass index (BMI)) on the performance of the model(s) will be explored. Approach. Whole-body PET scans of 42 patients (41 18F-FDG and one 68Ga-PSMA) scanned with standard radiotracer dosage were included in this study. One 18F-FDG patient data was set aside and the remaining 40 patients were split into four subsets of 10 patients with different mean patient BMI. Multiple copies of a developed cycle-GAN network were trained on each subset to predict standard scan images using 1/8 and 1/16 short duration scans. Also, the models’ performance was tested on a patient scanned with the 68Ga-PSMA radiotracer. Quantitative performance was tested using peak signal-to-noise ratio (PSNR), structural similarity index (SSIM), and normalized root mean squared error (NRMSE) metrics, and two nuclear medicine specialists analyzed images qualitatively. Main results. The developed cycle-GAN model improved the PSNR, SSIM, and NRMSE of the 1/8 and 1/16 short scan duration inputs both 18F-FDG and 68Ga-PSMA radiotracers. Although, quantitatively PSNR, SSIM, and NRMSE of the 1/16 scan duration level were improved more than 1/8 counterparts, however, the later were qualitatively more appealing. SUVmean and SUVmax of the generated images were also indicative of the improvements. The cycle-GAN model was much more capable in terms of image quality improvements and speed than the NLM denoising method. All results proved statistically significant using the paired-sample T-Test statistical test (p-value < 0.05). Significance. Our suggested approach based on cycle-GAN could improve image quality of the 1/8 and 1/16 short scan-duration inputs through noise reduction both quantitively (PSNR, SSIM, NRMSE, SUVmean, and SUVmax) and qualitatively (contrast, noise, and diagnostic capability) to the level comparable to the standard scan-duration counterparts. The cycle-GAN model(s) had a similar performance on the 68Ga-PSMA to the 18F-FDG images and could improve the images qualitatively and quantitatively but requires more extensive study. Overall, images predicted from 1/8 short scan-duration inputs had the upper hand compared with 1/16 short scan-duration inputs.

To assess the impact of respiratory phase matching between single-photon-emission computed tomography myocardial perfusion imaging (SPECT-MPI) and low-dose computed tomography (CT) for attenuation correction (AC). Forty patients underwent 1-day 99mTc-tetrofosmin pharmacological stress/rest SPECT-MPI using a cadmium-zinc-telluride gamma camera. Low-dose CT for AC was performed at deep-inspiration breath-hold. SPECT-MPI was acquired once with free-breathing (FB) and repeated at deep-inspiration breath-hold (BH) to match the respiratory phase of AC. From these acquisitions we reconstructed four data sets: free-breathing SPECT-MPI without AC (non-corrected; FB-NC), breath-hold SPECT-MPI without AC (non-corrected; BH-NC), free-breathing SPECT-MPI with AC (FB-AC), and breath-hold SPECT-MPI with AC (BH-AC), the latter representing respiratory-phase-matched AC SPECT-MPI. We compared semi-quantitative segmental tracer uptake, visual diagnosis, inter-observer agreement, and image quality. Compared with FB-NC, deep-inspiration BH-NC increases inferior and lateral uptake, but decreases septal uptake. Addition of AC to FB increases inferior and septal uptake, but decreases anterolateral uptake. Combining breath-hold MPI with breath-hold CT AC (BH-AC) increases inferior, inferolateral, and septal uptake, but reduces apical uptake, without affecting anterolateral uptake, with significant differences to all other protocols. Frequency of normal scans increases across protocols: 10% with FB-NC, 21% with BH-NC, 38% with FB-AC, and 51% with BH-AC. Image quality and inter-observer agreement were highest for BH-AC among all protocols. Compared with non-corrected breath-hold SPECT-MPI and with free-breathing AC SPECT-MPI, respiratory-phase-matched AC SPECT-MPI significantly affects segmental semi-quantitative uptake, increases the frequency of normal scans, yields the best inter-observer agreement, and significantly improves image quality. These findings suggest a potential role of respiratory triggered SPECT-MPI in clinical routine.

To obtain high-quality pre-treatment localization MR (sMR) images from dynamic cine-MR using the Swin-ResViT network for target tracking in MRgRT. We propose a ResViT model fused with a Swin Transformer module (Swin-ResViT) with an optimized bottleneck layer structure for enhancing feature extraction efficiency. Seventeen liver cancer patients were retrospectively enrolled from Sun Yat-sen University Cancer Center from February to July 2024, and 12 of them were assigned to the training set (using intra-treatment cine-MR and pre-treatment planning MR), with the remaining 5 patients as the test set. Image generation quality and model performance were comprehensively evaluated by quantifying the normalized root mean square error (NRMSE), peak signal-to-noise ratio (PSNR), structural similarity index (SSIM), motion marker point error, and model inference speed between sMR and reference localization MR. Regarding image quality, Swin-ResViT reduced NRMSE and LPIPS by 90% and 82% compared to cine-MR (P<0.001), and improved PSNR, SSIM, and CNR by 157%, 79%, and 181% (P<0.001), respectively. Regarding structural accuracy, the mean localization error of motion markers at the right hepatophrenic junction in the generated dynamic sMR sequences was 0.7695±0.7294 mm (P<0.05). Regarding model inference speed, for a single 224×224-pixel frame, the average processing time on an NVIDIA GeForce RTX 2080 Ti GPU was 15.5 ms for Swin-ResViT as compared with 41.4 ms for the ResViT network, demonstrating a 62% reduction. The Swin-ResViT model can synthesize high-quality sMR from cine-MR images. This method combines computational efficiency with significant image enhancement advantages, and thus has important clinical significance for real-time MRgRT.

Recent compressed sensing (CS) approaches utilize the similarity and redundancy of image patches of the magnetic resonance (MR) image to enable reconstruction from highly undersampled k-space measurements. In this paper, the patches similarity and redundancy is exploited by assuming that the image patches sample a low-dimensional manifold embedded in a high dimensional space. The MR image is then reconstructed by keeping the dimension of the patch manifold as small as possible. This is achieved by using the dimension of the patch manifold of a MR image as a regularizer in the objective function. Results show that the proposed method is superior to state-of-the-art patch-based MRI reconstruction methods. The algorithm is evaluated on two datasets containing 100 MR images each. The reconstruction quality of the algorithm, gauged using the three quality metrics: peak signal-to-noise ratio (PSNR), structural similarity index measure (SSIM) and normalized root mean square error (NRMSE), is better than the comparison methods.

Conventional whole-body (WB) PET imaging using short axial field-of-view (SAFOV) scanners is constrained by low sensitivity and limited temporal resolution, particularly for dynamic and kinetic parametric imaging. This study introduces a supervised deep learning framework to enhance WB Patlak parametric images (Ki) from SAFOV scanners by leveraging high-sensitivity data from long axial field-of-view (LAFOV) uEXPLORER system. Dynamic <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">${ }^{18}$</tex> F-FDG PET scans were acquired from two cohorts: 15 subjects using the uEXPLORER and 19 subjects using the Siemens Biograph mCT (SAFOV) with a 6 -minute single-bed acquisition in the cardiac region followed by 13 WB passes via continuous bed motion. Low-quality (LQ) WB parametric images were simulated from LAFOV data using <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$15 \sec \times 13$</tex> frames with 4 -minute interval (short-frame protocol), while how-quality (HQ) parametric images were reconstructed from LAFOV standard protocol. The image data starting from 20 minutes were used to reconstruct Patlak images. A 2D conditional consistency model (denoted as cDDPM-CM) was developed, incorporating LQ Ki images and prior PET images (average of dynamic frames converted to SUV) as conditioning inputs. The model was trained via a 5 -fold cross-validation on LAFOV data (4680 slices). Results demonstrated that the cDDPM-CM significantly outperformed traditional denoising methods (Gaussian filtering, non-local means) and a baseline DDPM-CM model (without prior PET image) in enhancing LQ LAFOV images, achieving a peak signal-to-noise ratio (PSNR) of 39.08, normalized root mean square error (NRMSE) of 0.3445, and structural similarity index measure (SSIM) of 0.9433. Similar improvements were observed in SAFOV images, with enhanced clarity and reduced noise.

Speech compression plays a significant role in speech communication systems. It yields a compact data representation allowing an efficient storage and transmission of information. The potentiality of speech compression has been essential to the technologies of robust long-distance communication, high-quality speech storage, and message encryption. Storage and archival of large volumes of spoken information makes speech compression essential and which improves the capacity of communications relatively of unlimited bandwidth. The main objective of this research work is to present a good quality speech data at a low bit rate. In order to accomplish this, the most powerful speech analysis and compression techniques such as Linear Predictive Coding (LPC), Discrete Cosine Transformation (DCT) and Discrete Wavelet Transformation (DWT) are adopted for Tamil speech database. The adopted techniques are evaluated based on Compression ratio, Peak Signal to Noise Ratio (PSNR), Normalized Root Mean Square Error (NRMSE) and Mean Opinion Score (MOS). Among these methods the DWT achieves greater performance than other two techniques employed in this research work.

In recent years, data security has been a challenging endeavor, especially when the data is being transmitted every second. Internet of Things (IoT) involves continuously sending data over public networks, making the data vulnerable to various security threats. Therefore, ensuring the secure end-to-end communication of IoT data is critical. Cryptography and steganography have proven effective in providing secure connectivity for IoT devices. However, challenges in existing approaches include scalability, computational complexity, implementation, key management, trade-offs, evolving threats, and hyperparameter tuning. Therefore, in this paper, we propose EGCrypto, an efficient and secure model for IoT networks. EGCrypto utilizes a low-complexity elliptic Galois cryptography approach along with matrix XOR steganography to enhance security. To optimize its performance, we employ the zoning evolution of control attributes and adaptive mutation based self-adaptive differential evolution with fitness and diversity ranking. These techniques are utilized to fine-tune the hyperparameters of EGCrypto, enhancing its effectiveness and efficiency. The confidential IoT data is encrypted using low-complexity elliptic Galois cryptography. Following encryption, the encrypted data is embedded or hidden into cover blocks of an image, which are selected using the optimization algorithm. This ensures secure data communication in IoT architectures, as the encrypted data is transferred safely and can be easily recovered and decrypted at the receiving end. The experimental results demonstrate that EGCrypto outperforms competitive models with improvements of 1.8473% in peak signal to noise ratio (PSNR), 1.5490% in strutural similarity index metric (SSIM), 1.7682% in normalized root mean square error (NRMSE), 1.3829% in carrier capacity, and 1.9372% in embedding efficiency.

Ischemic heart disease remains a leading cause of mortality worldwide. Myocardial perfusion imaging (MPI) using Rubidium-82 (82Rb) positron emission tomography (PET) is a cornerstone in its evaluation. However, conventional CT-based attenuation correction (AC) is prone to artifacts, with misalignment between PET emission data and the CT-AC being a common problem. This study evaluates the feasibility of introducing a deep learning approach to generate synthetic CT (sCT) images directly from non-attenuation-corrected 82Rb-PET images. To this end, we developed a cGAN using a conditional generative adversarial network (cGAN) with an Attention U-Net generator to produce sCT images to produce sCT-AC maps, based upon 544 PET/CT MPI scans. Image quality was assessed using structural similarity index (SSIM), peak signal-to-noise ratio (PSNR), mean absolute error (MAE), and mean error (ME). Additionally, attenuation-corrected PET images based on sCT were evaluated in the cardiac region using relative mean error (RME) and relative mean absolute error (RMAE). Cardiac function and perfusion assessments, defined as the ischemic total perfusion deficit (iTPD) and the left ventricular ejection fraction reserve (LVEFR), were compared between sCT-based and conventional CT AC methods. Our sCT-images provided good correlation to the conventional CT-AC (SSIM = 0.91 ± 0.037, PSNR = 29.9 ± 3.2dB). For the PET images, we report a slight bias in the cardiac region (RME = 4.2 ± 7.8%, RMAE = 6.9 ± 5.9%), likely due to a uniform overestimation of the soft-tissue u-maps within the sCT. Despite the bias, the quantification metrics remained comparable to those obtained with CT AC (mean iTPD: CT 3.73 ± 5.19% vs. sCT 3.67 ± 5.13%; mean LVEFR: CT 5.88 ± 5.96% vs. sCT 5.90 ± 6.11%). Additionally, the sCT-based approach appeared to reduce motion and implant-related artifacts, providing further motivation for its use over CT. This observation was made through visual inspection on a case-by-case basis. These results demonstrate the potential of deep learning-based sCT generation to maintain integrity in PET MPI while helping to mitigate issues related to misalignments and metal-induced artifacts.

The accurate and rapid prediction of hazard gas dispersion fields in urban environments is essential for guiding emergency sensor deployment and enabling real-time risk assessment. However, the computational cost associated with Computational Fluid Dynamics (CFD) simulations hinders their use as real-time forward models, while simplified Gaussian plume models lack the fidelity to resolve building obstruction effects. This study proposes a morphology-guided conditional Generative Adversarial Network (cGAN) framework designed to achieve real-time gas dispersion field modeling in urban environments with complex building configurations. The urban area is discretized into 50 × 50 m grid cells, each characterized by six morphological parameters describing building geometry. K-means clustering categorizes these cells into distinct morphological types. High-fidelity dispersion datasets are then generated for each type using Lattice Boltzmann Method (LBM) simulations. Each sample encodes building geometry, release location, wind speed, and time as multi-channel input images, with the corresponding gas dispersion concentration field is recorded as the output. Two cGAN architectures, Image-to-Image Translation (Pix2Pix) and its high-resolution variant (Pix2PixHD), are employed to learn the mapping from input features to dispersion fields. Model performance is evaluated using four complementary metrics: Fraction within a Factor of Two (FAC2) for prediction accuracy, Normalized Root Mean Square Error (NRMSE) for precision, Fractional Bias (FB) for systematic error, and Structural Similarity Index (SSIM) for spatial pattern fidelity. A case study is conducted across a 1176 km2 urban district in China. The results demonstrate that under varying wind speeds (0.5-1.5 m/s) and temporal scales (5-60 s), and across five morphological categories, the Pix2PixHD-based model achieves 92.5% prediction accuracy and reproduces 97.6% of the spatial patterns. The proposed framework accelerates computation by approximately 18,000 times compared to traditional CFD, reducing inference time to under 0.1 s per scenario. This sub-second capability enables real-time concentration field estimation for emergency management, and provides a physically informed, computationally feasible forward model that can potentially support sensor-based gas source localization and detection network planning in complex urban environments.

Electroencephalographic (EEG) recordings are often contaminated with artefacts, such as eye blinks, which complicate their analysis. While various methods exist to address, identify, and mitigate artefacts, many require human intervention. This study introduces a novel, self-supervised, fully automated approach for identifying and reducing artefacts in EEG signals using a Variational Autoencoder (VAE) architecture. In detail, subject-specific VAEs, with convolutional layers, are trained from spatially preserved EEG topographic maps. A sample-wise strategy based on the negative log-likelihood of activated latent vectors from training data is proposed to identify anomalous topomaps. This assigns an anomaly score to each model’s input. The vectors of input topomaps above a chosen threshold are automatically clipped with a percentile-based approach of activated latent space components. Eventually, the reconstructed EEG signals are compared with a baseline built upon an offline ICA method with automatic detection of artefactual components inspired by the FASTER methodology. Results show that the signal-to-noise ratio (SNR) and the peak signal-to-noise ratio (PSNR) of the FP1, FP2, and other channels were higher, while the remaining channels were similar to ICA Fast. Similarly, mean absolute error (MAE), normalised root mean square error (NRMSE), and correlation coefficients indicated comparable signals from both methods. In addition, findings demonstrate the method’s strength in avoiding signal updates in non-artefactual segments, preserving their neural dynamics. The contribution to the body of knowledge is a fully automated, subject-specific method for identifying and denoising EEG signals.

In super-resolution magnetic resonance imaging (SR-MRI), the low-resolution scans are acquired keeping the central low-frequency components intact. The scan-time for a given matrix size is shortened by acquiring only the central low-frequency part of the k-space (Fourier space), and filling the unacquired portion with zeroes. While transforming the zero-padded low-resolution acquired k-space to the image domain through inverse Fourier transformation, an inherent blur and high-frequency oscillatory artifacts are manifested in the reconstructed image. State-of-the-art methods including reconstruction-based and deep-learning-based approaches have addressed this problem to a large extent; however, it remains a challenging problem to completely eliminate the remnant effects of blur or ringing in some of the subtle clinical features. In the proposed SR method, we have implemented an interleaved hybrid domain convolutional neural network (CNN) model consisting of a k-space and spatial domain network together with local residual connections. This helps to improve the accuracy of hyper-parameter estimation and effectively reduce the loss function, with the resulting advantage of attaining improved peak-signal-to-noise-ratio (PSNR), structural similarity index measure (SSIM) and normalized root mean square error (NRMSE) as compared to the state-of-the-art methods.

Functional electrical stimulation (FES) is widely used in neurorehabilitation to improve patients’ motion ability. It has been verified to promote neural remodeling and relearning, during which FES has to produce an accurate movement to obtain a good efficacy. Therefore, many studies have focused on the relationship between FES parameters and the generated movements. However, most of the relationships have been established in static contractions, which leads to an unsatisfactory result when applied to dynamic conditions. Therefore, this study proposed a FES control strategy based on the surface electromyography (sEMG) and kinematic information during dynamic contractions. The pulse width (PW) of FES was determined by a direct transfer function (DTF) with sEMG features and joint angles as the input. The DTF was established by combing the polynomial transfer functions of sEMG and joint torque and the polynomial transfer functions of joint torque and FES. Moreover, the PW of two FES channels was set based on the muscle synergy ratio obtained through sEMG. A total of six healthy right-handed subjects were recruited in this experiment to verify the validity of the strategy. The PW of FES applied to the left arm was evaluated based on the sEMG of the right extensor carpi radialis (ECR) and the right wrist angle. The coefficient of determination (R2) and the normalized root mean square error (NRMSE) of FES-included and voluntary wrist angles and torques were used to verify the performance of the strategy. The result showed that this study achieved a high accuracy (R2 = 0.965 and NRMSE = 0.047) of joint angle and a good accuracy (R2 = 0.701 and NRMSE = 0.241) of joint torque reproduction during dynamic movements. Moreover, the DTF in real-time FES system also had a nice performance of joint angle fitting (R2 = 0.940 and NRMSE = 0.071) and joint torque fitting (R2 = 0.607 and NRMSE = 0.303). It is concluded that the proposed strategy is able to generate proper FES parameters based on sEMG and kinematic information for dynamic movement reproduction and can be used in a real-time FES system combined with bilateral movements for better rehabilitation.

Remote sensing datasets offer a unique opportunity to observe spatial and temporal trends in water quality indicators (WQIs), such as chlorophyll-a, salinity, and turbidity, across various aquatic ecosystems. In this study, we used available in situ WQI measurements (chlorophyll-a: 17, salinity: 478, and turbidity: 173) along with Landsat-8 surface reflectance data to examine the capability of empirical and machine learning (ML) models in retrieving these indicators over Matagorda Bay, Texas, between 2014 and 2023. We employed 36 empirical models to retrieve chlorophyll-a (12 models), salinity (2 models), and turbidity (22 models) and 4 ML families—deep neural network (DNN), distributed random forest, gradient boosting machine, and generalized linear model—to retrieve salinity and turbidity. We used the Nash–Sutcliffe efficiency coefficient (NSE), correlation coefficient (r), and normalized root mean square error (NRMSE) to assess the performance of empirical and ML models. The results indicate that (1) the empirical models displayed minimal effectiveness when applied over Matagorda Bay without calibration; (2) once calibrated over Matagorda Bay, the performance of the empirical models experienced significant improvements (chlorophyll-a—NRMSE: 0.91 ± 0.03, r: 0.94 ± 0.04, NSE: 0.89 ± 0.06; salinity—NRMSE: 0.24 ± 0, r: 0.24 ± 0, NSE: 0.06 ± 0; turbidity—NRMSE: 0.15 ± 0.10, r: 0.13 ± 0.09, NSE: 0.03 ± 0.03); (3) ML models outperformed calibrated empirical models when used to retrieve turbidity and salinity, and (4) the DNN family outperformed all other ML families when used to retrieve salinity (NRMSE: 0.87 ± 0.09, r: 0.49 ± 0.09, NSE: 0.23 ± 0.12) and turbidity (NRMSE: 0.63± 0.11, r: 0.79 ± 0.11, NSE: 0.60 ± 0.20). The developed approach provides a reference context, a structured framework, and valuable insights for using empirical and ML models and Landsat-8 data to retrieve WQIs over aquatic ecosystems. The modeled WQI data could be used to expand the footprint of in situ observations and improve current efforts to conserve, enhance, and restore important habitats in aquatic ecosystems.

The accuracy and speed of a single image super-resolution using a convolutional neural network is often a problem in improving finer texture details when using large enhancement factors. Some recent studies have focused on minimal mean square error, resulting in a high peak signal to noise ratio. Generally, although the peak signal to noise ratio has a high value, the output image is less detailed. This shows that the determination of super-resolution is not optimal. Conditional Generative Adversarial Network based on Boundary Equilibrium Generative Adversarial Network, by combining Mean Square Error Loss and GAN Loss as a loss function to optimize the super-resolution model and produce super-resolution images. Also, the generator network is designed with skip connection architecture to increase convergence speed and strengthen feature distribution. Image quality value parameters used in this study are Peak Signal to Noise Ratio (PSNR) and Structural Similarity Index (SSIM). The results showed the highest image quality values using dataset validation were 26.55 for PSNR values and 0.93 for SSIM values. The highest image quality values using the testing dataset are 24.56 for the PSNR value and 0.91 for the SSIM value.