Research

Neurodevelopmental conditions, such as Autism Spectrum Disorder (ASD) and Attention Deficit Hy- peractivity Disorder (ADHD), present unique challenges due to overlapping symptoms, which make accurate diagnosis and targeted intervention difficult. Recent advancements in deep learning models have shown promise in diagnosing neurodevelopmental disorders. Several studies have reported sig- nificant improvements in disorder classification performance through more sophisticated graph neural network designs and have identfied salient features that may serve as potential biomarkers. Higher- order features that are critical to correctly determining neuro-developmental disorders are frequently missed by the majority of current methods, which rely on spatial convolutionfiltering to extract local in- formation from resting-state fMRI data. In this study, we provide a multi-modal fusion framework that combines three distinct kinds of data from each sample: graph-based representations, multiple types of correlation matrices, and BOLD signals from diffeerent regions of interest. We created three specific model architectures to process these modalities: a Kolmogorov-inspired graph convolutional network (KANGCNET) for graph-structured data, a residual network for evaluating correlation matrices, and an LSTM network for learning temporal properties from BOLD signals..................................................

Electrocardiography (ECG) is vital for diagnosing cardiovascular diseases, yet accurate interpretation remains difficult due to waveform complexity and limited well-annotated datasets. We introduce CARDxnosis, a medical knowledge–driven deep learning framework that integrates anatomical lead grouping and attention mechanisms to capture clinically meaningful relationships among ECG leads. The architecture features two encoders: one learns spatio-temporal patterns from raw ECG.........[Under review]

Deep learning has significantly propelled the performance of ECG arrhythmia classification, yet its clinical adoption remains hindered by challenges in interpretability and deployment on resourceconstrained edge devices. To bridge this gap, we propose EXGnet, a novel and reliable ECG arrhythmia classification network tailored for single-lead signals, specifically designed to balance high accuracy, explainability, and edge compatibility. EXGnet integrates explainable artificial intelligence (XAI) supervision during training via a normalized cross-correlation based loss, directing the model’s attention to clinically relevant ECG regions, similar to a cardiologist’s focus. This supervision is driven by automatically generated ground truth, derived through an innovative heart rate variability-based approach, without the need for manual annotation. To enhance classification accuracy without compromising deployment simplicity, we incorporate quantitative ECG features during training. These enrich the model with multi-domain knowledge but are excluded during inference, keeping the model lightweight for edge deployment. Additionally, we introduce an innovative multiresolution block to efficiently capture both short- and long-term signal features while maintaining computational efficiency. Rigorous evaluation on the Chapman and Ningbo benchmark datasets validates the supremacy of EXGnet, which achieves average five-fold accuracies of 98.762% and 96.932%, and F1-scores of 97.910% and 95.527%, respectively. Comprehensive ablation studies and both quantitative and qualitative interpretability assessment confirm that the XAI guidance is pivotal, demonstrably enhancing the model’s focus and trustworthiness. Overall, EXGnet sets a new benchmark by combining high-performance arrhythmia classification with interpretability, paving the way for more trustworthy and accessible portable ECG based health monitoring systems.

Efficiently recognizing modulated signals has grown more crucial as communication technologies advances. Unfortunately, obstacles including a large number of data and inaccurate identification provide a barrier to advancement in this field. Automatic modulation classification (AMC) based on deep learning (DL) has been suggested as a productive way to get good classification performance. However, under different noise situations, particularly at low signal-to-noise ratios (SNR), the majority of existing DL-AMC algorithms have limited generalization capabilities. In this paper, we developed three AMC models: a Bidirectional Long Short-Term Memory-Gated Recurrent Unit (BiLSTM-GRU), a hybrid Convolutional Neural Network (BiLSTM-GRU-CNN) for spatiotemporal features, and Bidirectional LSTM-GRU-CNN-Attention (AMC-BLGCA) with multihead attention. Our models can be implemented on the receiver end of a transmission system as modulation classifiers, which are capable of differentiating modulated schemes in various scenarios. According to the experiment, the AMC-BLGCA has the best overall accuracy on both datasets at 62.54% and 65.00% on RadioML2016.10a and RadioML2016.10b dataset respectively. The BiLSTM-GRU model takes second place with much fewer parameters and a very close accuracy score to the AMC-BLGCA model. The parallel hybrid- BiLSTM-GRU-CNN model has the lowest performance among the proposed methods, but still outperforms the recent AMC models. Our proposed methods are very lightweight and easily implementable to edge devices as they use fewer parameters and the need for computation resources is very low. The performance is 3% to 9% better than the current state-of-the-art (SOTA) AMC methods.

A common neurodegenerative disease, Alzheimer’s disease requires a precise diagnosis and efficient treatment, particularly in light of escalating healthcare expenses and the expanding use of artificial intelligence in medical diagnostics. Many recent studies shows that the combination of brain Magnetic Resonance Imaging (MRI) and deep neural networks have achieved promising results for diagnosing AD. Using deep convolutional neural networks, this paper introduces a novel deep learning architecture that incorporates multiresidual blocks, specialized spatial attention blocks, grouped query attention, and multi-head attention. The study assessed the model’s performance on four publicly accessible datasets and concentrated on identifying binary and multiclass issues across various categories. This paper also takes into account of the explainability of AD’s progression and compared with state-of-the-art methods namely Gradient Class Activation Mapping (GradCAM), Score-CAM, Faster Score-CAM, and XGRADCAM. Our methodology consistently outperforms current approaches, achieving 99.66% accuracy in 4-class classification, 99.63% in 3-class classification, and 100% in binary classification using Kaggle datasets. For Open Access Series of Imaging Studies (OASIS) datasets the accuracies are 99.92%, 99.90%, and 99.95% respectively. The Alzheimer’s Disease Neuroimaging Initiative-1 (ADNI-1) dataset was used for experiments in three planes (axial, sagittal, and coronal) and a combination of all planes. The study achieved accuracies of 99.08% for axis, 99.85% for sagittal, 99.5% for coronal, and 99.17% for all axis, and 97.79% and 8.60% respectively for ADNI-2. The network’s ability to retrieve important information from MRI images is demonstrated by its excellent accuracy in categorizing AD stages.

Nowadays, Cancer’s devastating impact is growing, taking thousands of lives prematurely each day. Lung cancer stands at the forefront of this grim reality. Timely and accurate cancer diagnosis is crucial, as it directly correlates with effective treatment and improved patient outcomes. In this paper, we proposed an ensemble deep-learning method for detecting and classifying lung cancers that greatly impact the Computer Aided Diagnosis (CAD) system. Initially, three deep convolutional neural networks (CNN) Transfer Learning Approaches, MobileNetV2, VGG19, and Resnet50, were used individually to perform classification. Then, these models are combined to perform better in lung cancer diagnosis using the fusion of chest CT and PET-CT images. This approach leverages the strengths of MobileNetV2, VGG19, and ResNet50’s pretrained weights for feature extraction, and then the extracted features are concatenated and used for classification through the weighted average ensemble technique. After an extensive experimental analysis, the proposed ensemble model achieved a test accuracy of 98.93%, which is better than the individual model performance (98.67% in MobileNetV2, 98.20% in VGG19, and 97.67% in ResNet50). It can be an efficient diagnostic tool for lung cancer detection, as the prediction results of the proposed deep learning model outperform the recent Transfer Learning approaches.

The classification of White Blood Cells (WBCs) is crucial for diagnosing diseases, monitoring treatment effectiveness, and understanding how the immune system functions. In this paper, we propose a deep learning approach to classify WBCs using Super Resolution Generative Adversarial Network (SRGAN) and Visual Geometry Group 19 (VGG19). Firstly, microscopic images of WBCs are generated using the SRGAN to obtain more precise and high-resolution images, which are then classified with a pretrained VGG19 classifier. Low-resolution (LR) images are inputted into the generator of SRGAN, and its discriminator compares the High-resolution (HR) image with LR, generating super-resolution images to minimize misclassification risks. A large dataset of 12,447 images containing four classes of WBCs (Eosinophil, Lymphocyte, Monocyte, and Neutrophil) is utilized to train and validate our proposed model. Following extensive experimental analysis, our proposed model achieves a test accuracy of 94.87 %, surpassing traditional Convolutional Neural Network (CNN), Recurrent Neural Network (RNN), Hybrid CNN-RNN models, and other conventional approaches. The generated images of SRGAN overcome challenges associated with misclassification due to the poor resolution of microscopic images, while the use of a pretrained model as a classifier reduces classification complexity.

Diabetes is a global epidemic that demands early diagnosis to mitigate its progression and associated complications. By avoiding complicated medical procedures, machine learning presents a potential approach for the precise and efficient identification of diabetes. This study introduces a novel two-layer stacked ensemble model for early-stage diabetes classification.We employed four rigorous feature selection methods to determine the most salient features. Three heterogeneous base learners were then trained using the extracted features, namely Light Gradient Boosting Machine (LGBM), Extreme Gradient Boosting (XGBM), and Categorical Boosting (CatBoost). Finally, these learners were integrated into a stacked ensemble architecture with SVM as the meta-learner. Our proposed model outperformed the existing state-of-the-art models and those reported in previous studies. The model achieved an impressive accuracy of 99.04%, along with a precision of 0.99, a recall of 0.98, an F1-score of 0.975, and a perfect AUC value of 1. These results highlight the effectiveness of the stacked ensemble model in accomplishing highly accurate diabetes prediction, paving the way for improved early intervention and disease management.

Social networks are ubiquitous, with billions of users worldwide. These channels are essential for communication and information sharing in modern life. Due to their popularity, fraudsters have targeted online platforms using fraudulent URLs. Traditional URL-classification methods, which use post-access features or static analysis are not able to keep up with hackers’ ever-changing strategies and lack the granularity needed for precise URL categorization. To address this issue, we have proposed a unique approach to detect malicious URL. We have developed an ensemble machine learning approach comprising of Decision Tree, Random Forest, Extra Tree, and XGboost along with feature selection with the help of Particle Swarm Optimization (PSO). Feature selection refines model performance and reduce computation complexity. In this working scheme, we have implemented different machine learning and deep learning techniques to evaluate and compare the methods’ accuracy and effectiveness with the proposed scheme. We have utilized ISCX-URL2016 dataset and achieved an accuracy of 98.46% by the proposed ensemble machine learning. By this experiment, we have observed that our proposed machine learning model outperforms deep learning techniques and the existing methods as well. Additionally, we have analyzed the runtime and figured out that our proposed scheme computes effectively within 56.98s as total execution time with the selected features.

Wireless networks for large-scale data transmission have raised security and privacy concerns in recent years. Thus, intrusion detection systems (IDS) and other precautions have been taken. Computer and network systems need intrusion detection technologies, yet many IDS are ineffective. The increase of feature space also affects the accuracy of machine learningbased IDS methods. Continuous attempts have been made to improve IDS efficacy and dependability. Recently, numerous deep learning methods have been used to improve efficiency. They usually lack feature selection knowledge, resulting in less efficient systems. In this study, a thorough comparative analysis of various feature selection methods have performed in order to identify the most effective set of features to improve the model performance. We analyzed the effectiveness of a variety of feature selection techniques, such as filter, wrapper, and embedded methods in extracting the most informative features from the benchmark NSL-KDD dataset. Subsequently, the selected features were passed to a hybrid deep learning model which incorporated with Convolutional Neural Network (CNN), advanced Recurrent Neural network (RNN) and fully connected Deep Neural Network (DNN) architecture. Our findings show that feature selection method significantly affects model performance. We show that some feature selection strategies regularly provide improved feature subsets, improving model accuracy and robustness by carefully studying their interaction with deep learning algorithms. It is experimented in this study that the selected features by XGBoost shows optimal accuracy of 99.38% in the chosen deep learning method.