Profit Forecasting for Daily Pharmaceutical Sales Using Traditional, Shallow, and Deep Neural Networks: A Case Study from Sabha City, Libya

Nail diseases, including such common conditions as fungus, and more serious issues like melanoma, may be important clues to the overall health and require a clear diagnosis to be treated. The purpose of the paper is to create a nail disease detection system based on the advanced machine learning methods, including transfer learning and federated learning. The research seeks to show how machine learning and federated learning can be combined to detect nail disease performance with high accuracy without having to share data. The data include pictures of diverse nail conditions including Acral Lentiginous Melanoma, Onychogryphosis, and Pitting among others that are checked to maintain the quality of data in a uniform manner to facilitate the effective training of the models. The most common feature extraction models are ResNet152V2, DenseNet201, MobileNetV2, and InceptionResNetV2 that produce between 1,280 and 2,048 features per image. These characteristics are then pooled to create a unified feature space of 6,784 dimensions which is further narrowed to five major characteristics with Linear Discriminant Analysis (LDA) to create an efficient form of classification. A range of classification models, including Deep Neural Networks (DNN) and Long Short-Term Memory (LSTM) and Bidirectional LSTM (Bi-LSTM) are compared, with the last one reaching the highest classification accuracy of 91.8%. The federated learning strategy enables the joint training of DL models by different clients to ensure data-privacy and has validation-accuracy rates exceeding 99-percent in both uniformly random and structured data distributions. The proposed federated learning-based models resulted high in both uniformly random and structured data distributions.

As deep neural networks (DNNs) become integral to critical applications, protecting their intellectual property (IP) has become paramount. Neural network watermarking is a technique that embeds unique identifiers into models, asserting ownership and deterring unauthorized use. However, sophisticated attacks can deactivate or remove these watermarks without significantly compromising model performance, undermining current protection strategies. In this paper, we introduce the first method for reactivating deactivated neural network watermarks in altered DNN models without requiring access to the original model parameters or training data. By formulating the reactivation process as an optimization problem, we employ projected gradient descent to identify new trigger inputs that restore the embedded watermark. Regularization techniques are incorporated to ensure these triggers resemble legitimate inputs, enhancing both stealth and practicality. Through experiments on various benchmark datasets and model architectures, we demonstrate the effectiveness of our method against common model alterations, including fine-tuning, pruning, and surrogate model attacks. Our work addresses a critical gap in DNN IP protection, offering a robust and practical solution for watermark reactivation. This empowers model owners to assert their rights even in the face of advanced adversarial tactics.

The expansion of the Internet of Things (IoT) into consumer applications demands robust and energy-efficient communication protocols. Long-range wide area network (LoRaWAN) is a key enabler, but its performance depends on optimal spreading factor (SF) allocation, where traditional adaptive data rate (ADR) mechanisms are inadequate in dynamic environments. This study presents a novel lightweight stacked-ML approach for spreading factor (LSML-SF) allocation in mobile IoT LoRaWAN network. We propose a stacked ensemble model that jointly combines a linear stochastic gradient descent classifier (log-loss), a gradient boosting model, and a deep neural network (DNN) through a logistic regression meta-learner. The LSML-SF is trained on a vast dataset of 225,109 samples generated from ns-3 simulations, and our model achieves an out-of-fold cross-validation accuracy of 85%. Importantly, we demonstrate the practical feasibility of our approach through a rigorous computational analysis, showing the DNN component requires only 12,602 parameters and 12.3k MAC operations per inference. When integrated into ns-3 simulations, our LSML-SF framework significantly outperforms traditional ADR mechanisms and existing ML approaches, improving the packet success ratio and reducing energy consumption, thereby extending the operational lifespan of consumer IoT devices.

Acknowledging recent breakthroughs in the context of deep bio-inspired neural networks, several architectural deep network options have been deployed to create intelligent systems. The foundations of convolutional neural networks are influenced by hierarchical processing in the visual cortex. The graph neural networks mimic the communication of biological neurons. Considering these two computation methods, a novel deep ensemble network is used to propose a bio-inspired deep graph network for creating an intelligent supply chain model. An automated smart supply chain helps to create a more agile, resilient and sustainable system. Improving the sustainability of the network plays a key role in the efficiency of the supply chain’s performance. The proposed bio-inspired Chebyshev ensemble graph network (Ch-EGN) is hybrid learning for creating an intelligent supply chain. The functionality of the proposed deep network is assessed on two different databases including SupplyGraph and DataCo for risk administration, enhancing supply chain sustainability, identifying hidden risks and increasing the supply chain’s transparency. An average accuracy of 98.95% is obtained using the proposed network for automatic delivery status prediction. The performance metrics regarding multi-class categorization scenarios of the intelligent supply chain confirm the efficiency of the proposed bio-inspired approach for sustainability and risk management.

Probabilistic sampling and deep neural encoder-decoder network for advancing scene text recognition on irregular datasets

Automated Identification of Mango Leaf Diseases Using Deep Convolutional Neural Networks

Abstract We introduce a new approach for constructing immersive virtual spaces by generating comprehensive 3D voxelised models that encompass both geometric and semantic scene representations from a single 360 $${}^{\circ}$$ RGB-D input. The proposed approach utilises a deep convolutional neural network for semantic scene completion (SSC), allowing the estimation of complete semantics and geometries of the scene. We design MDBNet a dual head model that simultaneously processes RGB and depth data using a perspective camera. Depth information is encoded using a flipped transcribed signed distance function (F-TSDF), capturing essential geometric shape characteristics. We extend the inference capabilities of MDBNet on RGB-D input of the perspective camera to accommodate 360 $${}^{\circ}$$ RGB-D by proposing MDBNet360. We employ RGB spherical-to-cubic projection and 3D rotation for depth point clouds, allowing for virtual reality (VR) space design with comprehensive spatial coverage. To our knowledge, this is the first work to extend a pre-trained SSC model, originally using perspective camera RGB-D input, to infer a 3D model from 360 $${}^{\circ }$$ RGB-D input. To assess acoustic properties, we measure parameters such as early decay time (EDT) and reverberation time (RT60) using the exponential sine sweep method (ESS). We used Unity with the Steam Audio plug-in for conducting simulations in virtual space. The proposed framework demonstrates better virtual space reconstruction and immersive sound generation, advancing semantically rich and spatially accurate virtual environments compared to the state-of-the-art (SOTA). Code and rendered sounds are available on GitHub: https://github.com/MonaIA1/Repo360 .

Upper blepharoplasty is the most common cosmetic procedure in East Asia. A natural Asian double eyelid features specific crease characteristics. AI advancements, such as UNet and PointRend, enhance medical image segmentation, aiding in post-blepharoplasty evaluation. This study applies deep neural networks to analyze facial images, providing morphological parameters to assist surgeons in assessing outcomes and planning revisions. This study included 102 eyes from 51 patients seeking for revisional blepharoplasty and 100 eyes from 50 volunteers with inborn double eyelid. Standardized images and videos were collected. The deep learning-based image analysis automatically evaluated four eyelid morphological parameters, including pre-tarsal show, corneal visibility ratio, dynamic value, and crease depth. Analysis was done on the agreement between the automated measures and the manual measurements. The parameters of the patients' and volunteers' eyelids were compared. FACE-Q surveys were used to measure patient-reported esthetic outcomes. The intraclass correlation coefficients between manual measures and automated measurements of pre-tarsal show, corneal visibility ratio, and dynamic value were 0.973, 0.975, and 0.965. At the long-term follow-up, the pre-tarsal show and crease depth decreased significantly, whereas the corneal visibility ratio and dynamic value increased significantly. FACE-Q scores demonstrated a high level of patient satisfaction for facial appearance (87.6) and were negatively correlated with pre-tarsal show (r = - 0.814, p = 0.000). The deep neural network technique automatically measured the eyelid morphology with excellent precision and reproducibility, enabling an objective evaluation of the surgical outcomes for blepharoplasty. This journal requires that authors assign a level of evidence to each article. For a full description of these Evidence-Based Medicine ratings, please refer to the Table of Contents or the online Instructions to Authors www.springer.com/00266 .

ABSTRACT Insurers and reinsurers providing capacity to cyber insurance risks have now realised that current pricing models, though effective to date, do not accurately estimate an actuarially fair premium. Increased cyber risk exposure from connected devices, the volume of unstructured data, limited loss experience, and evolving risk complexity have contributed to the challenges of accurately modelling cyber risks. Most current models are based on reported or economic losses collected from secondary sources. The urgent need to develop a hybrid pricing model that integrates loss exposures and qualitative risk perceptions for cyber insurance policies is evident. This paper proposes a machine‐learning approach for modelling cyber risks using neural networks. We developed a model that accurately estimates the probability of loss for various cyber risks across industry segments. We developed a multilayer neural network model to predict the likelihood of cyber risk. We used a structural equation model to examine the aggregate effects of cyber risk on associated exposures. The outputs of both models can be used to estimate the organisation's financial liability and determine appropriate insurance coverage. Our findings show that system vulnerability, user awareness, and cyber risk mitigation significantly affect cyber risk exposure, and that the models' predictive ability is statistically significant. Furthermore, the results of these models were highly useful in building cyber risk resilience and developing actuarial pricing for the selected sectors or industries.

River water quality in monsoon-driven subtropical basins exhibits strong seasonal variability driven by hydroclimatic forcing and increasing anthropogenic pressure, posing challenges for reliable assessment and management. Despite advances in water quality modeling, most Water Quality Index (WQI) prediction frameworks require extensive sampling and lack interpretability, limiting rapid baseline assessment during critical periods. This study develops the first integrated Explainable Artificial Intelligence (XAI) framework combining Machine Learning (ML), Deep Learning (DL), and Physics-Informed Neural Networks (PINNs) to predict, interpret, and spatially characterize seasonal water quality dynamics in the Padma River Basin, Bangladesh. Forty-four surface water samples collected during winter and monsoon seasons were evaluated using WQI assessment, explainable modeling, probabilistic uncertainty analysis, and spatial regionalization. Results show that seasonal variability dominates over spatial variability (p < 0.0001), with winter low-flow conditions promoting solute concentration and localized degradation, while monsoon discharge drives basin-wide dilution and recovery. Model performance is strongly region-dependent: Deep Neural Networks achieve the highest accuracy in winter (R2 = 0.98; RMSE = 1.16), whereas Ridge Regression and Voting Ensemble models perform more robustly during the monsoon (R2 ≈ 0.97; RMSE ≈ 1.01). Explainable AI analysis identifies NO3- emerged as the dominant contaminant (24.0 ± 36.3mg/L winter, 47.5 ± 68.7mg/L monsoon, with isolated samples exceeding WHO limits), whereas pH and DO exhibit dual seasonal influences. PINN-based data augmentation improves model generalization under limited sampling while preserving hydrochemical consistency. Monte Carlo simulations quantify prediction uncertainty and reveal seasonal shifts in WQI probability distributions, while spatial autocorrelation analysis identifies localized winter degradation hotspots and widespread monsoon improvement. The proposed physics-informed and explainable AI framework enhances predictive reliability, interpretability, and decision relevance, offering a transferable approach for uncertainty-aware water quality assessment and adaptive management in monsoon-affected, data-limited river basins.

Manifold-based models offer a computationally efficient alternative to directly transporting the thermochemical state in computational simulations of turbulent reacting flows, projecting the high-dimensional thermochemical state-space onto a low-dimensional manifold. Recent efforts have yielded a manifold-based model applicable to multi-modal combustion, enabling reconstruction of the thermochemical state from solutions to two-dimensional manifold equations in mixture fraction and generalized progress variable that are parameterised by three scalar dissipation rates. In coarse-grained simulations such as Large Eddy Simulation (LES), closure of the multi-modal manifold equations and subfilter variances/covariance requires closure of three filtered scalar dissipation rates. The present work adopts a data-based approach, providing closure for the three filtered scalar dissipation rates via deep neural networks (DNNs). High-fidelity datasets corresponding to an autoigniting n-dodecane jet flame and a bluff body swirl-stabilized confined lifted spray flame of two aviation fuels (Jet-A and C1) with different ignition propensities are leveraged to generate training data that spans a diverse range of thermodynamic conditions and combustion modes, including low- and high-temperature ignition regimes in addition to premixed and nonpremixed behaviour. A final DNN model is trained to enforce inherent physical constraints by learning nonlinear functional transformations of the three filtered scalar dissipation rates. The generalizability of this constrained DNN model is demonstrated a priori via conditional statistics evaluated on the lifted spray flame with C1–a configuration that had not been included in the training data. Excellent DNN agreement with conditional DNS statistics is observed, and integrated gradients are computed to identify the most sensitive input variables. The similarity of the marginal PDFs of the most informative input variables and outputs across configurations are quantified via the Wasserstein metric, demonstrating that data-based models may successfully generalize to unseen parametric conditions so long as the most informative input variables share similar distributions across training and testing datasets.

A synergistic integration of physics-based and data-driven approaches has emerged as promising research field for terrestrial evapotranspiration (ET) estimation, enabling robust modeling of land-atmosphere interactions. This study proposes a hybrid model by integrating machine learning (ML)-based canopy surface resistance (rs,c) estimation into the Shuttleworth-Wallace (S-W) dual-source scheme under the ETMonitor framework, replacing traditional physics-based rs,c parameterization. Three ML algorithms, Random Forest (RF), Gradient Boosting Regression Tree (GBRT) and Deep Neural Network (DNN) were tested in the hybrid model. A reference dataset of rs,c was derived by inverting S-W dual-source model with in-situ flux measurements. The model was trained on 179 global flux tower sites and independently validated on 45 sites. Three full ML-based models based on DNN, GBRT and RF, were also developed to estimate ET directly for comparison. The DNN-integrated hybrid model outperformed the original physics-based model, with Kling-Gupta Efficiency (KGE) increasing from 0.7 to 0.84 and coefficient of determination (R²) increasing from 0.66 to 0.72. The three full ML models showed comparable performance to the hybrid models. Notably, the physics-ML hybrid framework balances physical interpretability with data-driven efficiency, minimizing reliance on prior knowledge and avoiding over-parameterization.

Quantification of interstitial fibrosis in digitized kidney biopsies using deep neural networks.

Monitoring the specific conductance (SC) in coastal zones is vital for environmental management and sustainable development. Due to unpredictable reasons such as atmospheric conditions, mechanical problems, power outages, sensors limits, etc., recording systems may fail which causes gaps in data recording. In this study, original artificial intelligence (AI) models are developed for the modeling and reconstruction of missing SC data. Two novel swarm-based deep neural networks (DNNs)-the nonlinear group method of data handling (NGMDH) and a long short-term memory (LSTM) model integrated with the turbulent flow of water-based optimization (TFWO) algorithm were developed and applied to model SC records. The results were also compared with six conventional and two ensemble machine learning (ML) models. The efficacies of the models were evaluated in five hypothetical scenarios. Then, in the derivation phase, the best models were applied to the SC datasets comprising 5% gaps. The results highlighted the extraordinary role of AI-based models in improving knowledge on SC distribution in coastal waters. The new LSTM-TFWO and NGMDH-TFWO models, with average normalized root mean square error (NRMSE) of 0.11 and 0.11, and R² of 0.742 and 0.71, are approximately 11% and 6.36% more accurate than LSTM and NGMDH models, respectively. However, the tree-based models, with an average NRMSE of 0.05, demonstrate substantially higher accuracy than these complex DNN architectures. Among all the ML methods evaluated, ensemble models showed superior performance in reconstructing gaps in SC datasets. XGBoost achieved the highest accuracy, as indicated by an NRMSE of 0.031. Consequently, ensemble models are recommended for application in simulating various types of engineering problems.

Abstract Minor changes in process variables, such as temperature, processing speed, and cooling rate, can significantly impact the properties of the final product in a sheet extrusion process. As a result, many optimization efforts focus on each of these variables. This study explored a machine learning-assisted process design for a polypropylene/carbon black composite sheet. The extrusion process parameters were selected as input variables, while tensile strength, void content, width, and thickness were measured, resulting in a dataset of 180 entries. A deep learning neural network was employed to identify and propose optimal combinations of process parameters and validate these proposals by comparing predicted values to experimental data. The polymer melt index had a significant effect on tensile strength, which is attributed to the degree of crystallinity. The process optimization led to a 20% increase in the tensile strength of continuous fiber composites, enhancing the matrix's toughness and improving interfacial load-carrying capacity.

The ability to quickly learn and generalize is one of the brain's most impressive feats and recreating it remains a major challenge for modern artificial intelligence research. One of the most mysterious one-shot learning abilities displayed by humans is one-shot perceptual learning, whereby a single viewing experience drastically alters visual perception in a long-lasting manner. Where in the brain one-shot perceptual learning occurs and what mechanisms support it remain enigmatic. Combining psychophysics, 7 T fMRI, and intracranial recordings, we identify thehigh-level visual cortex as the most likely neural substrate wherein neural plasticity supports one-shot perceptual learning. We further develop a deep neural network model incorporating top-down feedback into a vision transformer, which recapitulates and predicts human behavior. The prior knowledge learnt by this model is highly similar to the neural code in the human high-level visual cortex. These results reveal the neurocomputational mechanisms underlying one-shot perceptual learning in humans.

Accurate characterization of focal liver lesions on multiphasic magnetic resonance imaging (MRI) is central to early detection and treatment planning; however, clinical deployment of deep learning remains constrained by compute cost, latency, and the need for rigorous verification of hardware implementations. This paper presents LiverNet-Q, a hardware-aware, multi-phase deep neural network for multi-class liver disease detection from MRI, coupled with an end-to-end hardware co-simulation workflow that validates functional equivalence between the software model and the synthesized register-transfer level (RTL) design. The proposed pipeline first localizes the liver with a lightweight U-Net trained using public liver MRI annotations, then classifies lesions into five clinically relevant categories (normal liver, hepatocellular carcinoma, hemangioma, focal nodular hyperplasia, and simple cyst) using attention-based phase fusion. To enable resource-efficient inference, LiverNet-Q is trained with quantization-aware training and deployed in INT8 precision. The accelerator is implemented with Vitis HLS using a streaming dataflow micro-architecture that targets an initiation interval of one for core convolution operators. Experiments on public benchmarks demonstrate that INT8 deployment preserves diagnostic performance with a small loss relative to FP32 while providing substantial speedups. Hardware co-simulation reports confirm cycle-accurate latency and throughput, supporting reproducible, deployment-ready evaluation.

Abstract Metasurfaces capable of realizing complex electromagnetic responses provide new design routes to the microwave absorption. The traditional optimization process of metasurface is time-consuming and computational resource-consuming. With the help of advanced deep learning methods, the design of the metasurface can be accelerated. In this paper, the trained residual neural network model is combined with an adjusted genetic algorithm to effectively reduce the optimization time of the metasurface pattern and structural parameters and achieve a broadband effective absorption. The trained residual neural network has good prediction ability, with 99.48% of the samples of the test set having a loss value below 5×10-5; the adjusted genetic algorithm is better able to search the potential space where the metasurface pattern is located. The optimal metasurface absorber covers the frequency range from 7.83 – 18.0 GHz (EAB 10.17 GHz) with a thickness of 3.97 mm. The combination of deep neural networks and optimization algorithm provide an effective way for the fast design of metasurface absorbers.

This article introduces a distributed deep neural network (NN)-based adaptive control framework for cooperative object manipulation in human-robot teams with unknown agent dynamics by using three distinct multilayer NN observers (MNNOs). The first observer, termed the reference point estimator, enables each robotic agent to estimate the object's reference center using consensus-based learning, even without direct access to global reference trajectories. The second observer, referred to as the human force-to-trajectory estimator, uses human-applied forces to infer the intended position, velocity, and acceleration of the object, enabling real-time estimation of human intent. Together, these two observers allow distributed estimation of human-intended motion. In addition, a third observer, the distributed NN dynamics observer, is integrated into the control layer to simultaneously estimate the agent's own state and unknown system dynamics while incorporating the state vector of all other agents. Weight update laws for the multilayer NN observers are developed using singular value decomposition (SVD), enabling stable and efficient parameter tuning in multiagent settings. The framework combines the observer estimates with a distributed online multilayer actor-critic NN controller to compute Pareto game theoretic optimal effort that coordinates robot actions while considering neighborhood interactions. Safety is enforced via barrier Lyapunov functions (BLFs) formulated using Karush-Kuhn-Tucker (KKT) conditions, which dynamically adjust safety constraints based on both the agent's own state and its neighbor state vector. Simulation results demonstrate that the proposed approach achieves accurate intent estimation, robust control, and a 60% reduction in total cost compared to baseline methods.