Suhaila Abuowaida

Intrusion detectors for Internet of Things (IoT)-enabled substations must run on gateways with limited memory, yet the strongest published detectors are heavy tree ensembles. This study compresses a five-member soft-voting ensemble into compact multilayer perceptron (MLP) and gradient-boosted decision tree (GBDT) students on Distributed Network Protocol 3 (DNP3) flows (six attack families) and on phasor measurement unit (PMU) records of the MSU/ORNL dataset (attack versus natural events). Students were trained with class-weighted cross-entropy, weighting-only controls, and standard, minority-aware (MA-KD) and balanced distillation, and evaluated under stratified, chronological, and capture-held-out protocols. A class-weighted GBDT matched the teacher on both layers (macro-F1 0.986 versus 0.987 on DNP3; 0.908 versus 0.903 on PMU data) with 2.4% and 1.1% of its resident memory and 43.8 and 72.8 times lower single-thread CPU time per record, and stayed within 0.007 macro-F1 of the teacher under the stricter protocols. Distillation did not improve GBDT students. For MLP students, MA-KD was best on the stratified PMU split (0.734 versus 0.718 for standard distillation), whereas on DNP3 square-root class weighting without a teacher was best. Under the stricter protocols, PMU macro-F1 fell to about 0.60–0.63 for the teacher and all compact students (ExtraTrees: 0.53–0.55), identifying generalisation, not model size, as the main open problem.

Super-resolution (SR) facilitates high-fidelity mapping beyond the spatial resolution capacity of clinical neuro-oncology MRI. While state-of-the-art deep learning algorithms have augmented this domain, such methods exclusively learn from spatial or spectral domains and are often compounded by tumor boundary over-smoothing and geometrically inconsistent pathologies. In this paper, we introduce Hy-DTGL, a hybrid-domain transformer with a global–local attention modal super-resolution (SR) framework that learns a joint reconstruction across four inter-dependent knowledge domains: local spatial, global spatial, spectral and wavelet. Using windowed self-attention, Hy-DTGL learns highly localized pathologic textures; pooled global attention constrains anatomical plausibility; Fourier-learnable modulation provides spectral energy per semblance; and multi-scale discrete wavelet tiling preserves multi-scale tumor morphology. Optimization occurs via a hybrid objective, combining Charbonnier loss, gradient edge-loss, and spectral domain loss. The proposed framework was trained on BraTS 2019 (BT_19) and evaluated on BT_19, BraTS 2020 (BT_20), BraTS 2023 (BT_23), and IXI under both in-domain and strict zero-shot settings, and achieves competitive performance relative to the evaluated baselines under the adopted degradation and evaluation protocol at ×2, ×4, and ×8 SR factors. In the BraTS 2020 ×4 test resolution, 37.46 dB PSNR and 0.986 SSIM are attained, outperforming deep learning CNN-based SR (ECFNet) by 2.47 dB and transformer-based (WavTrans) by 4.05 dB in PSNR. Ablative studies reveal the spatial, spectral and multi-scale constraints notably converge to a faithfully mapped image space with restored tumor boundary delineation and global-structure consistency. We also show the Hy-DTGL generalizability by how it improves downstream tumor segmentation metrics: the super-resolved scans improve the segmentation Dice coefficient by 0.063 and significant reductions are observed in the Hausdorff Distance from 19.05 mm to 13.01 mm relative to the original low-resolution images, suggesting potential utility for downstream neuro-oncological image analysis.

One of the major challenges in applying machine learning (ML) to leukaemia diagnosis is the limited interpretability of deep learning models such as artificial neural networks (ANNs) and convolutional neural networks (CNNs). While these models excel at detecting complex patterns in medical data, they often function as “black boxes,” providing little transparency into their decision-making processes. This lack of explainability raises concerns in clinical practice, where understanding why a model predicts a specific leukaemia subtype is critical for guiding treatment decisions, including chemotherapy and bone marrow transplantation. To address these concerns, researchers are increasingly adopting explainable AI (XAI) techniques. Methods such as saliency maps, feature importance rankings, and attention mechanisms allow clinicians to visualize how specific factors such as genetic mutations, biomarkers, or cell morphology contribute to diagnostic outcomes. These interpretability tools not only enhance transparency but also foster trust among healthcare providers and patients. By bridging the gap between computational accuracy and clinical reliability, XAI ensures that ML models can be safely integrated into diagnostic workflows. Ultimately, explainable approaches strengthen confidence in AI-driven leukaemia diagnosis, supporting more informed medical decisions and improving patient care.

The super-resolution (SR) of Magnetic Resonance Imaging (MRI) is essential for utilizing clinical scans with limited resolution, noise, and anisotropic sampling, such as multi-modal brain tumor imaging. In this work, we propose a Tri-Representation hybrid framework for MRI SR, TRT-GLA, that redefines the MRI SR task as a joint spatial–spectral–structural high-resolution image generation problem. TRT-GLA utilizes (i) spatial global–local attentions for modeling the spatial anatomy, (ii) a Fourier spectral transfer mechanism for upholding spectral consistency, and (iii) multi-scale hierarchical spectral decomposition for improved edge details. To adapt the learning framework to medical imaging characteristics, we introduce a tri-representation consistent loss function that explicitly combines pixel-wise, spectral, and edge structure priors from the high-resolution ground-truth, as well as a progressive resolution learning strategy. Our large-scale brain tumor experiments, on the IXI, BraTS 2019, 2020, and 2023 datasets, show that TRT-GLA achieves state-of-the-art results at upsampling factors of ×2, ×4, and ×8, respectively, achieving substantial improvements across CNN, GAN, and transformer-based methods in Peak Signal-to-Noise Ratio (PSNR), Structural Similarity Index (SSIM), and Multi-scale Structural Similarity Index (MS-SSIM). We further demonstrate how SR benefits brain tumor segmentation through the downstream task evaluation of a dual-branch segmentation framework. TRT-GLA produces highly accurate tumor segmentation results from low-resolution inputs, improving over native high-resolution inputs at ×8 in critical tumor boundary regions and in small tumor regions. There remains a small gap between native, high-resolution imaging and SR-enhanced performance, which TRT-GLA nearly closes under realistic scenarios. Our results highlight the importance of synthesizing unified priors over spatial, spectral, and structural domains within a transformer for anatomically faithful reconstructions. Importantly, we also establish the utility of TRT-GLA in supporting quantitative analysis through a downstream tumor segmentation experiment that is clinically relevant.

The purpose of this research was to create a fully automatic deep-learning method to accurately classify different types of cardiac arrhythmias based on patterns observed in electrocardiograms (ECGs). There are several challenges that arise when manually interpreting cardiac arrhythmias. These include variability in how different observers interpret arrhythmias and variability in the diagnostic results obtained by different clinicians. To address these issues, we propose an architecture that includes both residual transfer learning and transformer encoder blocks. In our proposed architecture, the residual learning block uses Conv1D layers with skip connections to learn hierarchical representations of features in the ECG signal. The transformer block uses multi-head self-attention to identify longer-range dependencies in the ECG sequence. Our proposed model is tested on two publicly available benchmark databases of ECG recordings, namely the MIT-BIH Arrhythmia Database and the PTBDB Diagnostic ECG Database. We evaluate the performance of our model using a stratified 10-fold cross-validation procedure as well as Receiver Operating Characteristic (ROC) analysis. The proposed model achieved a classification accuracy of 99% on both datasets, with precision scores of 0.88 to 0.99 and recall scores of 0.87 to 0.99 for each arrhythmia class. Furthermore, the AUC values of the ROC analysis ranged from 0.98 to 1.0, indicating high levels of discrimination against minority classes, including supraventricular ectopic beats and fusion beats.

The heterogeneity of breast tissue and subtle morphological variations in ultrasound images make breast cancer detection a challenging task. This study proposes a hybrid deep learning framework that integrates EfficientNetB4 and ConvNeXt within a dual-stream architecture enhanced by advanced attention mechanisms. The model combines multi-scale texture representation with spatial feature extraction to improve classification performance. A two-stage preprocessing pipeline, consisting of adaptive median filtering and bilateral filtering, is applied to reduce speckle noise while preserving important structural details. The proposed method is evaluated on BUSI and UDAIT datasets, achieving 87.82% accuracy, 87.33% precision, and 85.33% recall on BUSI, and 85.69% accuracy, 84.00% precision, and 78.00% recall on UDAIT. These results outperform several baseline models, including ResNet-50, DenseNet-121, and vision transformers. Error analysis shows limitations in detecting small lesions and cross-modal generalization, with reduced performance on mammography images. Attention visualization demonstrates strong agreement with radiologist annotations, supporting model interpretability. The findings highlight the effectiveness of hybrid architectures for ultrasound-based breast cancer detection while emphasizing the need for modality-specific optimization.

Early and accurate breast cancer diagnosis remains critical yet challenging in routine practice. This study proposes a simple, reproducible pipeline that combines deep feature extraction from pre-trained CNNs (ResNet50, VGG16, EfficientNet-B0, DenseNet121, MobileNetV2) with classical machine-learning classifiers (logistic regression, SVM, k-NN, decision tree, random forest, gradient boosting, XGBoost, LightGBM, Naïve Bayes, and MLP). Features are computed after standardized preprocessing; class imbalance is addressed with SMOTE when present. We evaluate three image datasets (binary and multiclass) using accuracy, precision, recall/sensitivity, F1, and confusion matrices, and apply paired statistical tests across cross-validation splits. Findings: EfficientNet-B0+MLP and ResNet50+MLP achieve peak accuracies up to 99.6% on high-quality, balanced data, while DenseNet121+MLP with SMOTE attains 97.8% on imbalanced multiclass data. SMOTE yields substantial gains on imbalanced data and negligible effect on balanced sets; decision trees underperform consistently. Novelty/Improvement: Rather than a monolithic end-to-end network, we provide a modular, resource-aware blueprint that (i) disentangles feature extraction from classification, (ii) quantifies when imbalance correction matters, and (iii) reports clinically relevant error types. We further outline explainability with Grad-CAM/SHAP and discuss inference-time trade-offs and real-world workflow integration, offering an interpretable and deployment-friendly alternative to heavier end-to-end models.

Early skin disease detection significantly improves patient survival rates, yet limited access to dermatological expertise creates an urgent need for automated diagnostic systems. In this paper, we develop a dual-branch deep learning framework that simultaneously performs skin lesion segmentation and classification from dermoscopic images. The proposed segmentation branch uses a modified EfficientNet-B7 encoder with Atrous Spatial Pyramid Pooling (ASPP) for multi-scale feature extraction and transformer blocks for global context modeling. Attention gates and Squeeze-and-Excitation blocks enhance feature selection and boundary precision. The classification branch fuses DenseNet-121 visual features with morphological characteristics extracted from predicted segmentation masks, creating a hybrid appearance-morphology analysis approach. The proposed framework achieved strong and consistent segmentation performance across five benchmark datasets. On HAM10000, the highest Dice score (0.9568) and IoU (0.9242) were recorded, with an accuracy of 0.9708. PH2 achieved a Dice of 0.9250 and a sensitivity of 0.9734, while ISIC 2016 reached a Dice of 0.9298 and an IoU of 0.8811. For ISIC 2017 and ISIC 2018, Dice scores were 0.8972 and 0.9020, respectively. All datasets reported high specificity (> 0.93) and accuracy (> 0.95), confirming the model’s robustness and generalization capability. Our dual-branch framework achieves state-of-the-art accuracy by effectively integrating visual appearance and structural morphological features for comprehensive skin lesion analysis. The consistent high performance across diverse datasets indicates strong potential for clinical deployment as a diagnostic support tool.

Alzheimer's Disease (AD) is a chronic neurodegenerative disease that affects a large portion of the global population, and early and accurate diagnosis is the key component to proper management and treatment. This work presents the MAGNet model, a novel Deep Learning (DL) architecture for AD classification based on multi-modal imaging data. The MAGNet model uses multi-modal attention that is capable of hierarchically fusing structural MRI, functional MRI, and PET scans to obtain different kinds of information. The MAGNet model was tested on three datasets, with an overall accuracy of 96.2% when distinguishing between the AD, MCI, and CN groups. The proposed approach surpasses benchmark models by achieving 3.5% better accuracy and 5.2% higher sensitivity for early MCI diagnosis. Moreover, the MAGNet model offers interpretable results, employing attention visualization to support clinicians' decisions. MAGNet has the potential to predict cognitive scores and brain age with MMSE errors of 1.8 and a brain age of 2.3 years.