We introduce SAN++, an enhanced slicing adversarial network for compressed-sensing MRI reconstruction that integrates three key novelties: a transformer-guided attention block (TGAB), an edge-aware adaptive sampling module (EAASM), and a self-supervised pretraining strategy using masked image modelling (MIM). SAN++ extends prior GAN-based MRI methods by incorporating global transformer-based attention to better capture long-range dependencies, and by learning dynamic k-space sampling masks guided by salient image edges, which preserves critical structural features. We pretrain the network with a masked reconstruction task on large unlabeled datasets, then fine-tune adversarially with multi-resolution generators and a sliced optimal transport loss. Experiments on MRI datasets under various undersampling ratios (4 & times;, 8 & times;) and noise levels demonstrate that SAN++ outperforms DAGAN, RefineGAN. SAN++ achieves higher PSNR/SSIM and lower LPIPS perceptual error across settings, and shows robust performance under noise and sampling variability. An ablation study confirms each component's benefit, notably a 2-3 dB PSNR gain from TGAB and EAASM. Our results (with quantitative tables and example reconstructions) highlight the efficacy of combining transformer-guided attention, adaptive sampling, and self-supervised pretraining in adversarial MRI reconstruction.
Type 2 diabetes mellitus (T2D) in adolescents is a rising public health concern, yet early detection is challenged by limited and imbalanced datasets. This study compares traditional machine learning (ML) and deep learning (DL) methods for predicting adolescent prediabetes using a dataset of 209 records (11.5% prediabetic). Preprocessing included imputation, feature reduction, normalisation, and SMOTE-based class balancing. Baseline ML models - logistic regression, random forest, support vector machine (SVM), and multi-layer perceptron were evaluated via stratified cross-validation, with SVM achieving the best F1-score (0.955). To address class imbalance, synthetic data were generated using generative adversarial networks (GANs) and Wasserstein GANs (WGANs). A 1D convolutional neural network (CNN) trained on the augmented dataset achieved 98.5% accuracy, 96.0% precision, 97.0% recall, and a 96.5% F1-score on the original test set. Results confirm the value of GAN-based augmentation combined with CNNs for improving prediction under limited data, supporting timely T2D risk identification in adolescents. Unlike previous machine learning studies that relied solely on statistical resampling or small neural models, our approach combines GAN-based data synthesis with a one-dimensional convolutional network, yielding a substantial improvement in predictive power and generalisability under limited data conditions.
This research is designed to develop an advanced framework for the prediction of diabetes and the personalised recommendation of drugs. This research proposes a multi-attention enhanced task incremental learning coupled gradient boost deep neural network (MA-TIL-GBNN) approach aimed at the prediction of diabetes and recommendation of drugs based on their types. The approach integrates MA techniques into the DNN and generative adversarial network-based data augmentation (GDA) to effectively analyse the health indicators. Additionally, the TIL enhances the training of data, and the Light GBM facilitates efficient processing. The experimental results using the Diabetes Health Indicators Dataset showed 97.79% accuracy, 98.22% sensitivity, and 97.36% specificity, respectively.
Lie detection using an electroencephalogram (EEG) signal has received immense attention. A finite impulse response filter preprocesses the input EEG signal from the lie wave's datasets and the frequency split-up process. If a person lies, signal strength increases; if it exceeds a limit, a lie is detected. However, the convolutional methods produce robustness and false positive rates. The Tetralet attention-enabled modified N-Adam optimised distributed capsule network (Tet-MNDCNet) is proposed. A combination of Tetralet attention-enabled modified N-Adam optimised distributed capsule network and the zero-attention mechanism ensures Tetralet accuracy and reliability. The Tetralet attention is focused on selective features, enhancing relevant EEG patterns for improved lie detection accuracy. The Tet-MNDCNet model's performance is robust due to the HarmoniQ spectrum from the pre-processed signal. N-Adam optimiser reduces the gradient descent problem and improves the model's interpretability. The accuracy of the proposed experimental lie detection task of 97.35% for K-fold is 10.
Brain tumour classification is a significant research area in medical imaging. Manual examination of MRI scans is time-consuming, laborious, and may lead to imprecise findings. With the growth of artificial intelligence, automated methods are increasingly used for accurate detection of different brain tumour types. This paper presents a computer-aided diagnostic technique based on a 16-layer CNN architecture for precise tumour classification. The MR images are first resized and normalised, followed by dataset balancing using a hybrid SMOTE-edited nearest neighbour method. The balanced images are then fed into the proposed CNN model. A CNN-based feature extractor is also used with machine-learning classifiers including random forest, kNN, SVM, na & iuml;ve Bayes, and decision tree. Experimental results show that the proposed model achieves 98.88% accuracy for binary tumour detection and 97.83% for three-class classification, demonstrating its efficiency in detecting and identifying brain tumour types.
Epilepsy is a chronic neurological disorder that occurs due to irregular brain activities. An automated approach to detect the epileptic seizure state from EEG recordings is highly desirable as the manual approach is exhausting, time-consuming, and error-prone. This work presents a hybrid 1D-CNN + stacked-LSTM model for an end-to-end, patient-specific EEG-based epileptic seizure state detection. The proposed work was tested on two datasets: CHB-MIT scalp EEG dataset and Siena scalp EEG dataset. It achieved highest result of 97.07% accuracy, 97.80% sensitivity, 97.07% specificity, 0.0293 FPR, and 0.99 AUC values on CHB-MIT dataset and 97.83% accuracy, 98.75% sensitivity, 97.82% specificity, 0.0218 FPR, and 0.99 AUC values on Siena scalp EEG dataset. The results obtained were compared with latest patient-specific seizure state detection methods. The proposed model achieved best patient-specific results despite the challenges of varying channels, recording duration, and seizure intervals.
The steady-state visual evoked potential (SSVEP) brain-computer interface (BCI) has attracted widespread research interest owing to its multi-target recognition capacity, high accuracy, and efficient information transmission. However, the recognition accuracy of wearable SSVEP-BCI systems remains limited. To address this issue, this study proposes a feature transfer-based bidirectional long short-term memory (FTBi-LSTM) classification model, which incorporates variational mode decomposition (VMD) and wavelet hybrid denoising for signal preprocessing. Within the framework of bidirectional signal processing, SSVEP signals and same-frequency reference signals are paired as input for the bidirectional sub-networks. Deep features are extracted using a feature transfer approach to achieve classification. Experimental results show that under a 0.5-second time window, the classification accuracies for dry and wet electrodes reached 44.71% and 68.23%, while under a 0.2-second time window, the information transfer rates (ITR) increased to 142.96 bits/min and 337.42 bits/min, respectively, demonstrating the effectiveness of the FTBi-LSTM model in wearable SSVEP-BCI systems.
Real-time biomechanical head simulation is necessary for providing bio-feedbacks for facial paralysis grading. This process is challenging and needs enhancement in both dataset and personalising procedure. We introduced a statistical framework for dataset generation, skull prediction, and muscle strain computation. The head-to-skull shape relation was trained through their shape parameters. After a ten-fold cross-validation, the mean testing error was 1.86 mm with 6.17s +/- 0.05s for each fold. The personalised muscle network could be animated by interacting with the system interface for computing the muscle strains. This study has three contributions: a system for personalising and analysing biomechanical head; a procedure for head region cutting and sampling; head-and-skull shapes with their topological features. In perspective, this framework will be used to enhance the accuracy of the head-to-skull prediction. Moreover, we will use the system to generate the standard muscle strains for facial paralysis diagnosing. The dataset is available upon reasonable requests.
Accurate measurement and analysis of dynamic foot pressure are crucial for preventing complications associated with peripheral neuropathy, particularly in individuals with diabetes. This study focuses on optimising the calibration and sensitivity of an optical pressure sensor-based dynamic pedograph, designed and made in Bangladesh, to enhance its accuracy in assessing foot pressure distribution. The system employs total internal reflection in a transparent glass slab; foot-applied pressure disrupts light propagation, producing scattered light that is captured as greyscale intensity by a camera positioned beneath. Calibration was performed using a custom four-pad platform, establishing a linear relationship between applied pressure and pixel intensity (100-200 out of 255). Dedicated Java-based software enabled real-time analysis and precise correlation. Systematic tuning of camera parameters - gamma 100, gain 0, contrast 0 - enhanced sensitivity, linearity, and spatial uniformity, with spatial sensitivity variation of 8.2% and temporal variation of 1.92, indicating stable performance. The optimised pedograph generates high-resolution pressure maps, providing a cost-effective, reliable alternative to commercial systems, supporting improved plantar pressure assessment and diabetic foot care in clinical settings.
Bileaflet mechanical heart valves (BMHV) are clinically used to replace diseased heart valves. This study models the aortic root structure using medical imaging data and employs numerical simulations to analyse the hemodynamic characteristics of BMHV with different leaflet curvatures under pulsatile flow. Simulation results show that curved leaflets - by increasing the middle jet orifice area - improve the uniformity of three-jet flow distribution. Increasing leaflet curvature makes curved leaflets induce rotational fluid motion on their surface, causing more chaotic downstream vortex distributions but gradually reducing vortex intensity. With higher leaflet curvature, time-averaged wall shear stress (TAWSS) decreases, while oscillatory shear index (OSI) tends to increase in the ascending aorta. Regions with low wall shear stress (WSS) and high OSI usually have higher relative residence time (RRT), and transverse oscillatory shear index (OSItr) also decreases with increasing leaflet curvature. Thus, optimising mechanical valve leaflet design to improve hemodynamic performance can reduce postoperative complications.
Squats with lateral resistance are used in patellofemoral pain syndrome (PFPS) rehabilitation to target vastus medialis obliquus (VMO) activation, but evidence is inconsistent, and patellofemoral joint loading in these variations remains understudied. This study was to determine the differences in lower-limb muscle activation and patellofemoral joint loading during three lunge squats in PFPS individuals. Twenty-nine college athletes with PFPS performed three lunges: traditional lunge, hip-adduction lunge, and hip-abduction lunge. One-way repeated measures ANOVA was used to compare the variables of interest among the three lunges. The results demonstrated that hip-abduction lunge significantly increased the activation of the gluteus maximus and gluteus medius while concurrently reducing patellofemoral joint loading (descent phase: 2.39 +/- 0.72 N/kg; ascent phase: 2.48 +/- 0.61 N/kg). For PFPS individuals, the hip-adduction lunge may be more appropriate than the other two lunges when exercising the hip muscles and minimising patellofemoral joint loading.
This study aimed to develop a vibrating instrument to apply axial dynamic loading during wound healing with titanium-doped beta-tricalcium phosphate (Ti-beta-TCP) implants. Healing outcomes were compared in rabbits using clinical, radiographic, histological, oxytetracycline labelling, micro-CT, and SEM analyses. Bone defects were created in the femoral condyle of three animal groups: unfilled controls (Group I), Ti-beta-TCP implants without loading (Group II), and Ti-beta-TCP implants with dynamic loading (Group III). Over two months, no acute inflammatory reactions were observed. Implant degradation indicated new bone formation, most pronounced in Group III. Histology showed improved bony structures with Haversian canals and satisfactory tissue regeneration in loaded implants. Oxytetracycline labelling confirmed higher new bone deposition in Group III. Micro-CT showed enhanced bone regeneration via implant degradation, and SEM revealed nearly bridged bone-implant interfaces in loaded samples. Overall, dynamic loading improved bone healing and implant integration compared to static conditions.
The detection of blood vessels in the retina helps to identify various diseases such as diabetes and hypertension. Detection of vessels is a complex task facing specialists during the segmentation process especially children's blood vessels, which are thin. We proposed a deep learning model to do the semantic segmentation of blood vessels in general and the identification of vessels with very high accuracy, where we used short discrete wavelet transform to enhance the features extracted from the deep learning that we created to fit the waves. We applied different types of discrete waves with varying scaling within the model to accurately detect vessels. In addition, we used these waves on other models of DL used for vascular segmentation, where the yield improved significantly after these additions. The experiments on the Digital Retinal Images for Vessel Extraction (DRIVE) database were our model achieved the best results with test F1-score and accuracy of 0.9873, 0.9787, respectively.
