Mostafa Saleh

Background/Objectives: Deep learning models for diabetic retinopathy (DR) usually report a severity grade, and their attention maps explain that grade rather than individual lesions. We aimed to detect microaneurysms, hemorrhages, hard exudates and soft exudates (cotton-wool spots) separately, each with an attention map trained to fall on that lesion. Methods: An EfficientNet-B4 and Transformer encoder, pretrained on severity grades, feeds four lesion queries that each return a score and an attention map; a localization loss trains each map toward its lesion mask. We trained it on DDR and TJDR, tested it on their test sets and on two external sets (IDRiD and e-ophtha), and compared it with a U-Net. Results: The mean areas under the curve (AUCs) were 0.93–0.94 for hemorrhages, 0.90–0.96 for hard exudates and 0.81–0.90 for soft exudates; microaneurysms reached 0.91–0.96, except on DDR (0.58). We analyze this drop in detail. Attention ranked lesion pixels above the rest of the retina with an AUC-Judd of 0.948–0.996. Without the localization loss, detection was slightly lower, and attention moved to regions outside the lesions that still raised the prediction. The model detected lesions better than U-Net (mean AUC of 0.876 against 0.853) and, unlike the latter, also learned from images that have labels but no masks: with masks on half of the training images, detection and AUC-Judd were nearly unchanged. Conclusions: Supervising each lesion’s attention with its own mask gives a per-lesion score and a map that falls on that lesion, without reducing detection.

The threat of quantum computers to public-key cryptography has spurred a rapid search for post-quantum cryptography (PQC) standards, leading NIST to standardize ML-KEM (FIPS 203), ML-DSA (FIPS 204), and SLH-DSA (FIPS 205). Porting these algorithms to resource-constrained devices is challenging. A software-only approach is too slow even on optimized cores, while existing accelerators either consume a lot of area overhead or omit one of the two dominant kernels (NTT and Keccak), leaving the corresponding bottleneck in software. We present QUASAR, the smallest hardware–software co-design for the CV32E40P RISC-V core that accelerates all three NIST PQC algorithms simultaneously. Two loosely coupled engines (LUNA for NTT, SPOKE for Keccak), selected through a profiling-driven design-space exploration of coupling strategies, add only 886 LUTs—a +19% overhead over the bare processor core. On a mid-range Zynq-7000 SoC at 110 MHz, our proposed hardware–software co-design achieves speedups of 2.49–2.57 × (60–61% cycle reduction) for ML-KEM, 2.36–2.46 × (58–59%) for ML-DSA, and 3.10–3.18 × (68–69%) for SLH-DSA and a 3.21 × lower area–time product than the next-best design supporting both key encapsulation and digital signature algorithms.

This study evaluates advanced deep learning models — specifically CNNs (EfficientNetB2 and DenseNet201) and transformers (DeiT and ViT) — for the multi-label classification of chest X-ray images, with a primary focus on AI model development and optimization. Using the small CheXpert dataset for training/validation and the original CheXpert test set for evaluation, the research addresses all 14 labels (12 thoracic pathologies, “No Findings,” and “Support Device”), tackling the inherent challenges of overlapping pathologies in multi-label classification. While DeiT showed the highest performance, EfficientNetB2 was chosen for further refinement because it surpassed DenseNet201 on overall metrics and, along with DenseNet201, supports Grad-CAM for post-hoc explainability. Grad-CAM attention maps identified clinically relevant regions in correctly classified cases and pinpointed areas needing improvement in misclassified instances, thereby providing transparency into the models’ decision-making processes. Overall, despite the limitations imposed by lower image resolution and task complexity, the study demonstrates the adaptability of these deep learning models and contributes to the development of trustworthy AI systems in medical imaging that effectively balance diagnostic accuracy with interpretability. Our proposed method achieved promising results with 0.81 accuracy, 0.79 AUC, and integrates explainable AI by providing visual explanations through attention maps using Grad-CAM to enhance model interpretability.

Offensive-language detection systems that perform well at a given point in time often degrade as linguistic patterns evolve, particularly in dialectal Arabic social media, where new terms emerge and familiar expressions shift in meaning. This study investigates temporal linguistic drift in Saudi-dialect offensive-language detection through a systematic evaluation of continual-learning approaches. Building on the Saudi Offensive Dialect (SOD) dataset, we designed test scenarios incorporating newly introduced offensive terms, context-shifting expressions, and varying proportions of historical data to assess both adaptation and knowledge retention. Eight continual-learning configurations—Experience Replay (ER), Elastic Weight Consolidation (EWC), Low-Rank Adaptation (LoRA), and their combinations—were evaluated across five test scenarios. Results show that models without continual-learning experience a 13.4-percentage-point decline in F1-macro on evolved patterns. In our experiments, Experience Replay achieved a relatively favorable balance, maintaining 0.812 F1-macro on historical data and 0.976 on contemporary patterns (KR = −0.035; AG = +0.264), though with increased memory and training time. EWC showed moderate retention (KR = −0.052) with comparable adaptation (AG = +0.255). On the SimuReal test set—designed with realistic class imbalance and only 5% drift terms—ER achieved 0.842 and EWC achieved 0.833, compared to the original model’s 0.817, representing modest improvements under realistic conditions. LoRA-based methods showed lower adaptation in our experiments, likely reflecting the specific LoRA configuration used in this study. Further investigation with alternative settings is warranted.

Background: Chest X-ray (CXR) is widely used for the assessment of thoracic diseases, yet automated multi-label interpretation remains challenging due to subtle visual patterns, overlapping anatomical structures, and frequent co-occurrence of abnormalities. While recent deep learning models have shown strong performance, limitations in interpretability, anatomical awareness, and robustness continue to hinder their clinical adoption. Methods: The proposed framework employs a hybrid ConvNeXtV2–Vision Transformer (ViT) architecture that combines convolutional feature extraction for capturing fine-grained local patterns with transformer-based global reasoning to model long-range contextual dependencies. The model is trained exclusively using image-level annotations. In addition to classification, three complementary post hoc components are integrated to enhance model trust and interpretability. A segmentation-aware Gradient-weighted class activation mapping (Grad-CAM) module leverages CheXmask lung and heart segmentations to highlight anatomically relevant regions and quantify predictive evidence inside and outside the lungs. An ontology-driven neuro-symbolic reasoning layer translates Grad-CAM activations into structured, rule-based explanations aligned with clinical concepts such as “basal effusion” and “enlarged cardiac silhouette”. Furthermore, a lightweight out-of-distribution (OOD) detection module based on confidence scores, energy scores, and Mahalanobis distance scores is employed to identify inputs that deviate from the training distribution. Results: On the VinBigData test set, the model achieved a macro-AUROC of 0.9525 and a Micro AUROC of 0.9777 when trained solely with image-level annotations. External evaluation further demonstrated strong generalisation, yielding macro-AUROC scores of 0.9106 on NIH ChestXray14 and 0.8487 on CheXpert (frontal views). Both Grad-CAM visualisations and ontology-based reasoning remained coherent on unseen data, while the OOD module successfully flagged non-thoracic images. Conclusions: Overall, the proposed approach demonstrates that hybrid convolutional neural network (CNN)–vision transformer (ViT) architectures, combined with anatomy-aware explainability and symbolic reasoning, can support automated chest X-ray diagnosis in a manner that is accurate, transparent, and safety-aware.

Background/Objectives: The clinical adoption of AI in radiology requires models that balance high accuracy with interpretable, anatomically plausible reasoning. This study presents an integrated diagnostic framework that addresses this need by unifying a hybrid deep-learning architecture with explicit anatomical guidance and neuro-symbolic inference. Methods: The proposed system employs a dual-path model: an enhanced EfficientNetV2 backbone extracts hierarchical local features, whereas a refined Vision Transformer captures global contextual dependencies across the thoracic cavity. These representations are fused and critically disciplined through auxiliary segmentation supervision using CheXmask. This anchors the learned features to lung and cardiac anatomy, reducing reliance on spurious artifacts. This anatomical basis is fundamental to the interpretability pipeline. It confines Gradient-weighted Class Activation Mapping (Grad-CAM) visual explanations to clinically valid regions. Then, a novel neuro-symbolic reasoning layer is introduced. Using a fuzzy logic engine and radiological ontology, this module translates anatomically aligned neural activations into structured, human-readable diagnostic statements that explicitly articulate the model’s clinical rationale. Results: Evaluated on the NIH ChestX-ray14 dataset, the framework achieved a macro-AUROC of 0.9056 and a macro-accuracy of 93.9% across 14 pathologies, with outstanding performance on emphysema (0.9694), hernia (0.9711), and cardiomegaly (0.9589). The model’s generalizability was confirmed through external validation on the CheXpert dataset, yielding a macro-AUROC of 0.85. Conclusions: This study demonstrates a cohesive path toward clinically transparent and trustworthy AI by seamlessly integrating data-driven learning with anatomical knowledge and symbolic reasoning.

This paper introduces a lightweight, energy-efficient Number Theoretic Transform (NTT) implementation tailored for the ML-KEM scheme in post-quantum cryptography. Unlike prior efforts focusing mainly on multipliers and modular reduction, the proposed design emphasizes the address generator, a critical yet often overlooked component. A compact counter-based mechanism integrated with XOR logic and multiplexing eliminates large lookup tables, reduces redundant operations, and ensures efficient memory access for both NTT and INTT. The unified butterfly unit and a LUT-based modular reduction circuit enable multi-cycle modular operations without requiring digital signal processing (DSP) or block RAM (BRAM) resources. The complete architecture occupies only 147 slices and 557 LUTs, achieving very low power consumption. Implemented on a Virtex-7 platform, the design demonstrates up to 88.8% reduction in area, 22% decrease in dynamic power, and a 17.57× improvement in A²T efficiency compared with state-of-the-art implementations. These results highlight the significance of optimizing address generation to achieve compact, power-aware accelerators. The architecture’s elimination of memory-intensive components and reliance on LUT-based logic make it highly portable and suitable for resource-constrained devices.

Spectre variants 1 and 2 pose grave security threats to dynamic branch predictors in modern CPUs. While extensive research has focused on mitigating these attacks, little attention has been given to their energy and power implications. This study presents an empirical analysis of how compiler-based Spectre mitigation techniques influence energy consumption. We collect fine-grained energy readings from an HPC-class CPU via embedded sensors, allowing us to quantify the trade-offs between security and power efficiency. By utilizing a standard suite of microbenchmarks, we evaluate the impact of Spectre mitigations across three widely used compilers, comparing them to a no-mitigation baseline. The results show that energy consumption varies significantly depending on the compiler and workload characteristics. Loop unrolling influences power consumption by altering branch distribution, while speculative execution, when unrestricted, plays a role in conserving energy. Since Spectre mitigations inherently limit speculative execution, they should be applied selectively to vulnerable code patterns to optimize both security and power efficiency. Unlike previous studies that primarily focus on security effectiveness, this work uniquely evaluates the energy costs associated with Spectre mitigations at the compiler level, offering insights for power-efficient security strategies. Our findings underscore the importance of tailoring mitigation techniques to application needs, balancing performance, energy consumption, and security. The study provides practical recommendations for compiler developers to build more secure and energy-efficient software.

Pulmonary diseases rank among the leading causes of mortality worldwide, underscoring the importance of early detection to enhance patient outcomes and reduce fatalities. Chest X-rays (CXRs) serve as a critical diagnostic tool for identifying lung diseases; however, many pulmonary conditions exhibit similar patterns on CXRs, making differentiation challenging even for experienced radiologists. Furthermore, a single CXR may reveal multiple coexisting diseases, such as pneumonia and pleural effusion. Moreover, most existing studies in this field have been constrained to a limited number of classes and datasets, leaving significant gaps in the classification of diverse pulmonary conditions, particularly in remote regions where access to skilled radiologists is scarce. Recently, vision transformers (ViTs) have become a key technique in deep learning, and they have been widely used to detect different diseases from CXRs. This study aimed to adapt and optimize the pre-trained ViT-B16 model to predict and distinguish between single and dual pulmonary diseases whose diagnoses relied only on CXRs. The goal was to discriminate eight classes of evidence, which consisted of individual diseases (COVID-19, pneumonia, pneumothorax, tuberculosis, and pleural effusion), dual diseases (pneumonia and pleural effusion as well as pneumothorax and pleural effusion), and normal status. The experimental results show that the proposed model distinguished more classes than previous methods, with an accuracy of 98.18%. To the best of our knowledge, this is the first study to attempt to discriminate eight classes of evidence concerning pulmonary diseases, including both individual and dual diseases.

Spectre variants 1 and 2 pose grave security threats to the dynamic branch predictors in CPUs. This study examines experimentally the impact of mitigating those threats on energy and power consumption. The research obtained empirical data from an HPC-class CPU via embedded sensors with fine-grained energy readings. This approach black-boxed otherwise complex behaviors in and around the CPU. It was crucial to the concrete energy estimates reported here. The study identified factors that impacted the energy behaviors of branching (the attack vector for Spectre). It focused on compiler-based mitigation strategies vs. a no-mitigation baseline on three main compilers in wide use. It utilized a standard suite of microbenchmarks to gauge the consumption likely due to branch prediction under mixed branch and mitigation conditions. Results showed that the energy consumption varied significantly across the compilers, depending on the tasks. Loop unrolling affected the power consumption as it altered the number and distribution of branches. Speculative execution had a role in conserving energy. Reducing execution time was accompanied by an increase in instantaneous power. Since Spectre mitigations limit that feature, they should be applied selectively to the code patterns vulnerable to attacks. This work hopes to create an awareness of the energy costs of securing CPUs against Spectre attacks. It suggests directions for better energy-aware mitigation strategies. It highlights the importance of fitting compiler and optimization settings to application needs to balance security, performance, and power. The study provides compiler writers with recommendations that help build secure and power-efficient software.