Fusion: Practice and Applications

Kidney cancer is a lethal cancerous and very dangerous disease caused by genetic renal disease or by kidney tumors, and some patients might survive since there is no technique for earlier diagnosis of kidney tumor. Earlier diagnosis of kidney tumor assists physicians to begin proper treatment and therapy for the patient, which prevent kidney cancers and renal transplantation. Accurate classification of kidney tumor is vital for prediction and treatment planning. However, manual classification by pathologists could be subjective and time-consuming, and there can be considerable inter-observer variability. With the development of artificial intelligence (AI), automated tools enabled by machine learning (ML) and deep learning (DL) methods could predict cancers. This study designs a new white shark optimizer with an ensemble majority voting based kidney cancer classification (WSO-EMVKCC) technique on pathology images. The presented WSO-EMVKCC technique intends to identify the different grades of kidney cancer using DL and ensemble voting concepts. To accomplish this, the presented WSO-EMVKCC technique employs a deep convolutional neural network based Xception technique for the feature extraction process. Moreover, the WSO model has been used for the optimal hyperparameter tuning of the Xception approach. Furthermore, an ensemble majority voting classifier including three ML techniques like long short-term memory (LSTM), sparse autoencoder (SAE), and gated recurrent unit (GRU) models are employed for kidney cancer classification. The stimulation validation of the WSO-EMVKCC model is performed on the open access histology image database from Kaggle repository. The stimulated values illustrate the promising performance of the WSO-EMVKCC algorithm over other DL techniques.

The Internet of Things (IoT) devices and technologies are more effective in the medical sector. It includes the combination of numerous interrelated sensor, systems, and devices for gathering, examining, and conveying health-related information for medicinal uses. In the healthcare field, Blockchain (BC) technology embraces huge latent for increasing the security and confidentiality of data. BC-aided intrusion detection on IoT healthcare methods is a new technique for increasing the privacy and security of complex health data. Patients have superior control throughout their information’s growth, granting or revoking access as needed, but healthcare employees will modernize data sharing and certify the reliability of significant data. On the other hand, deep learning (DL) is excellent for transforming healthcare analytics, presenting rapid and tremendously precise estimations of medicinal circumstances. This paper presents a Blockchain-Assisted Deep Learning Model for Enhancing Healthcare Data Security with Metaheuristic Optimization Techniques (BCDL-HDSMOT) model. The main intention of the BCDL-HDSMOT technique is to develop an effective method for enhancing data security in the medical sector. At first, the blockchain technique is applied in healthcare to enhance data security, interoperability, and transparency while ensuring patient privacy and efficient record management. Next, the data pre-processing stage employs min-max normalization to clean, transform, and organize input data into a suitable quality for analysis. Besides, the black widow optimization algorithm (BWOA) has been deployed for the feature selection process to select the relevant features from input data. For the classification process, the proposed BCDL-HDSMOT technique designs a versatile long-short-term memory (VLSTM) method. At last, the improved seagull optimization algorithm (ISOA)--based hyperparameter selection process is performed to optimize the classification results of the VLSTM method. The experimental evaluation of the BCDL-HDSMOT algorithm can be tested on a benchmark dataset. The wide-ranging outcomes highlight the significant solution of the BCDL-HDSMOT approach to the cyberattack detection process.

In the digital age, e-commerce platforms are critical components of the global economy, facilitating seamless transactions and interactions between businesses and consumers. The digital infrastructure of these institutions is frequently attacked, either to hack or disrupt online services, leading to significant financial losses and damage to reputation. The most famous of these attacks are DDoS attacks, which lead to an increase in the volume of traffic to the platform's website beyond the capacity of the servers, thus causing the platform to respond slowly and crash and customers to be unable to access it. The increase in these attacks causes significant material damage to institutions, whether in the loss of revenues or the cost of responding to attacks. This work presents a robust DDoS attacks early detection model that can be adopted on e-commerce platforms using a lightweight one-dimension Convolutional neural network. The proposed model leverages the efficiency of deep learning with the lightweight architecture to analyze network traffic in real time, identifying patterns indicative of an impending DDoS attack. The balance between high detection accuracy with computational efficiency makes it suitable for real-time implementation in diverse e-commerce environments. DNN is trained on a comprehensive dataset of network traffic, encompassing both normal and attack scenarios, to ensure it can distinguish between legitimate traffic spikes and malicious activity. DDoS Evaluation Dataset CIC-DDoS2019 and CICIDS2017 are used in the experimental and accuracy achieved 0.98 and 0.99 in these two datasets respectively.

As increasingly Arabic textual data becomes accessible through the Intranet and Internet services, there is an important requirement for technologies and devices to handle the related data. Named Entity Recognition (NER) is an Information Extraction task that became a major part of several other Natural Language Processing (NLP) tasks. NER for Arabic has been obtaining improving attention, but possibilities for development in performance are even accessible. In recent decades, the Arabic NER (ANER) task has been confined to great effort to increase its performance. The ANER difficult task is to collect vast corpora or immense white gazetteers/lists that address probably the majority of Arabic language challenges like complexity, orthography, and ambiguity. Recently, deep learning (DL) has been the most typically applied NER model in the Arabic language and others. DL methods utilize the features of words and text to identify NEs. This paper presents a Secretary Bird Optimization Algorithm for Enhancing Fusion Deep Learning in Arabic Named Entity Recognition (SBOFDL-ANER) model. The main intention of the SBOFDL-ANER technique is to develop an effective method for NER in Arabic text. At first, the text pre-processing stage is applied to clean and transform the raw text into a structured format for analysis. Next, the word embedding method has been implemented by the Word2Vec method. Besides, the proposed SBOFDL-ANER technique designs ensemble models such as deep belief network (DBN), elman recurrent neural network (ERNN), and multi-graph convolutional networks (MGCN) for the process of classification. Eventually, the secretary bird optimization algorithm (SBOA) implements the hyperparameter choice of ensemble models. A wide-ranging simulation was applied to verify the performance of the SBOFDL-ANER method. The experimental outcomes demonstrated that the SBOFDL-ANER model highlighted improvement over other existing methods

Multi-label data stream classification plays a crucial role in various applications, including recommendation systems, real-time monitoring systems, smart cities, social media analysis, and healthcare. Its ability to classify constantly generated, potentially unbounded data at a high rate is of utmost importance. Besides accommodating multiple labels, data streams may evolve due to concept drift and bias toward particular classes due to class imbalance. This research introduces the multi-label classification model based on Hoeffding inequality (ML-kNN-H). The proposed model aims to process multi-label data streams, handle concept drift, and class imbalance. ML-kNN-H removes instances introducing errors based on a dynamic value computed from the Hoeffding inequality instead of a fixed value, thereby enhancing the model's efficiency and applicability to different types of data streams. Several experiments have been conducted to assess the model's performance in the presence of concept drift (abrupt and gradual drift) and class imbalance. Particularly, it has been evaluated against six kNN multi-label classifiers on ten datasets: synthetic and real world. The results indicate that ML-kNN-H outperformed the other classifiers on benchmark datasets in terms of Subset Accuracy, Accuracy, Hamming Score, and F-score, except in running time. Statistical analysis has also been utilized to measure the significance of the ML-kNN-H compared to the state-of-the-art classifiers.

Accurate disease prediction is essential for enabling preventive healthcare and reducing the burden of chronic illnesses. This study introduces an innovative multi-disease prediction framework leveraging machine learning and optimization techniques to address limitations in precision and scope present in prior research. Specifically, we focus on predicting five major diseases—diabetes, heart disease, kidney disease, liver disease, and breast cancer—by employing the Social Group Optimization (SGO) algorithm to fine-tune the Random Forest (RF) classifier's hyperparameters.The proposed SGO-optimized RF model optimizes seven critical parameters - n_estimators, max_depth, min_samples_split, min_samples_leaf, max_features, bootstrap, and criterion simultaneously, significantly enhancing the model's performance. Our approach, applied to five disease datasets, achieves notable accuracy improvements: 98.25 When tested on diverse datasets, the model achieves exceptional accuracy: 98.25% for breast cancer, 84.62% for liver disease, 93.44% for heart disease, 82.47% for diabetes, and 100% for chronic kidney disease. On average, the SGO-optimized RF outperforms existing methods with a 2.3% accuracy improvement across datasets. This research highlights the transformative potential of SGO-based optimization in advancing the accuracy and reliability of predictive models. The results demonstrate the framework's applicability in clinical scenarios, providing precise and actionable insights that support early diagnosis and improve patient outcomes.

The complex developing nature of urban infrastructure necessitates intelligent solutions for optimizing smart city operations. Based on this research paper, a multi-modal fusion framework that integrates real-time traffic and environmental sensor data with advanced machine learning algorithms to enhance decision-making for urban traffic management and pollution control is proposed. A hybrid AI model is proposed, with a combination of CNNs for the estimation of image-based traffic density, LSTM networks for the time-series environmental prediction, and RL for adaptive control of traffic signals. The system proposed integrates sensor data in real-time from cameras, GPS, LiDAR, and nodes for environmental monitoring to create an optimized control strategy. The model has been deployed on edge computing devices, such as Raspberry Pi, to enable the real-time processing and reduce the latency. Security layer based on block chain for data integrity protection and tamper proofing within smart city networks. The suggested system shows high improvements in congestion reduction, better accuracy in air pollution forecasting, and energy efficiency in urban management. It will be validated using simulation with SUMO and MATLAB and real-world sensor data that the sensor fusion approach outperforms the conventional fixed-rule strategies of traffic management. This work allows for cost-effective, large-scale smart city deployment that would reduce traffic delay and urban air pollution while securing data and being computationally efficient. The low-latency decision-making approach with edge-AI makes it fit for real-time urban governance. Unlike traditional models that process either traffic or environmental data in silos, the work presented herein integrates multi-source sensor data with edge computing and blockchain security for a unified AI-driven fusion approach, thus building a robust framework for next-generation smart city intelligence.

In medical diagnosis and prognosis, symptoms provided by patients play a critical role in identifying diseases. Machine learning offers a powerful approach to analyzing and predicting illnesses based on these symptoms. In particular, classification algorithms are widely used to analyze input data and predict disease outcomes. A key factor in effective classification is the selection of relevant attributes, which directly affects the accuracy of the prediction. This research emphasizes the importance of proper feature extraction techniques in the context of disease prediction using biomedical signal analysis. Effective analysis requires both the extraction of critical features and the elimination of irrelevant data. The aim of this study is to explore existing approaches to disease prediction based on biomedical signal analysis. We focus on feature extraction from pre-processed data, which aids in distinguishing between different biomedical signals recorded by medical devices. Our objective is to identify biomedical cues that differentiate various health conditions. Examples of such signals include electroencephalogram (EEG), electrocardiogram (ECG), and electrogastrogram (EGG). Understanding how these signals differ between healthy and diseased states is crucial for accurate disease prediction. This research investigates diseases such as heart disease, kidney failure, and lung infections, considering how variations in biomedical signals can be used to predict the likelihood of severe illness. We continue to seek advancements in predicting and mitigating future health risks

Precise assessment of software development effort (SDE) is essential for efficient project planning and resource distribution. Conventional methods frequently encounter difficulties in generalizing across different project areas because of disparate data attributes. This research presents an innovative approach that combines transfer learning with hybrid deep learning models to tackle these difficulties. The platform utilizes pre-trained Random Forest and LSTM models, enhanced using Jaya optimization, to improve prediction accuracy and adapt effectively to new datasets. Transfer learning is utilized to extract reusable patterns and features from source domains, facilitating effortless adaption to target domains with minimum retraining. Extensive experiments on various benchmark datasets illustrate the proposed framework's enhanced performance regarding accuracy, scalability, and robustness relative to leading techniques. This study emphasizes the capability of transfer learning to transform SDE estimates, providing a scalable and domain-adaptive approach for intricate software projects.

Predicting rainfall proves critical for businesses to organize their water resources, make agricultural choices, and prevent disasters. Therefore, proposed model presents a novel approach, namely Heuristic Intelligence towards Enhancing Rainfall Prediction with Artificial Neural Networks (HI2NN) to enhance rainfall prediction by designing heuristic Intelligence combined with Improved Artificial Neural Networks (IANNs). The proposed HI2NN framework leverages heuristic optimization techniques to fine-tune ANN parameters to improve prediction accuracy. Prediction accuracy is computed through our designed custom accuracy metric. The methodology uses historical weather information to extract complex non-linear patterns, which neural models generate from the designed big dataset. The accuracy level of rainfall predictions using our methodology achieves 92%, which demonstrates superior performance than traditional approaches that include random forest and decision tree and XGBoost models. The new forecasting systems develop higher reliability through collaborative efforts between heuristic algorithms and neural networks as described in this research work targeting challenging meteorological forecasts.

Plant diseases are considered a real threat to food security due to the losses incurred by individuals and countries. Early detection is one of the real solutions that can help reduce the size of these losses, but early detection is still bleeding. This study presents the development of a Convolutional Neural Network (CNN) model for classification with a new architecture and optimal performance suitable for real-time applications for the detection of fruit diseases (figs, oranges, grapes). The developed CNN model balanced accuracy and FLOPs using Squeeze-Excitation (SE) and adaptive-average pool layers. After implementing new data developed from Iraqi farms, the CNN model achieved optimal performance compared to the most famous models such as VGG16, ResNet, EfficientNet, and AlexNet.

Elderly health has always been a matter of concern for the medical doctors and researchers to come up with advanced recovery techniques. With the rise in population of elderly people and mostly residing alone at home in solitude has motivated many researchers to work on remedial measures for the biggest safety risk faced by them which is elderly fall prevention and mitigating thereby causes of injuries. In this paper, an intelligent deep learning and computer vision based elderly fall recognition system is designed which utilizes advanced spatial-channel decoupled downsampling in You Only Look Once version 10 (YOLOv10), pytorch, darknet and cascaded CNN technologies for the fall detection. The results after testing manifest that the accuracy of the proposed system to recognize and detect the elderly fall is quite assuring, the values of accuracy and mean Average Precision (mAP50) coming out to be 92.46% and 94.1% respectively after the model validation. Moreover, the system displays a real time performance as it can process approximately 45 frames of images per second that realizes a real-time identification of elderly fall patterns. As compared to previous models, the proposed model is much more efficient and has shown promising results.

One of the most important sectors for providing for daily human requirements is agriculture. At the same time, digitization has a big impact on a number of businesses, making it simpler to carry out a number of challenging tasks. In order to help the farmer and the consumer, technology and digitization must be adopted. Utilizing technology and routine monitoring, diseases can be identified and eliminated, increasing agricultural output. This paper suggests a system for recognizing and categorizing plant illnesses, initially focused on five separate classes: two fruit classes, one vegetable class, one edible pulse class, and one-grain class. The Plant Village and UCI ML Repository Dataset, which is well known as a freely accessible, accepted standard, and reliable data source, was used for this purpose. Based on them, a CNN model is prepared for analyzing them with an accuracy upto 95.42%. Image segmentation will also play a role in calculating precise amount of infection followingly, a good interface is must to utilize it in a proper way for a user which can be provided in the form of app, a feature that every user requires on daily basis.

Leukemia is a very dangerous kind of malignancy troubling the blood or bone marrow in all age categories, both in adults and children. The deadly and threatening kind of leukemia is named Acute Lymphoblastic Leukemia (ALL). The accurate and automated ALL diagnosis of blood cancer is complex work. Medical experts and hematologists in the bone marrow and blood samples detect it by employing a high-quality microscope. The manual classification is observed as tiresome and is restricted by varying expert considerations and other attributes. Presently, the Convolutional Neural Networks (CNNs) have become an acceptable mechanism for analyzing the medical image. However, for attaining outstanding performance, conventional CNNs normally demand large data sources for better training.   Thus, to alleviate the existing complexities, we implemented an effective ALL detection system using deep learning. At first, the necessitated images are aggregated from global resources of data. Further, the garnered images are inputted into the Optimized Trans-Res-Unet+ (OTRUnet+)-based segmentation model. Here, the Fitness-aided Position Updating in the Social engineering Algorithm (FPUSA) for improving the segmentation process’s efficacy optimally tunes the OTRUnet+ technique parameters.  In addition, the segmented images are taken to perform the classification process using the Adaptive Multi-Dilated Residual Attention Network (AMDRAN); here several parameters are optimally tuned by the same FPUSA to enrich the classification process. Finally, the suggested AMDRAN technique offered the ALL classified output. The effectiveness of the designed ALL detection system is explored with several existing systems to display its enhanced performance over other models

Identifying Autism early in children is vital for ensuring more precise developmental support and effective therapeutic interventions. Traditional diagnostic approaches are frequently delayed, and data privacy concerns limit the availability of broad, multi-institutional datasets required for effective machine learning models. To address these limitations, this study proposes a CNN-LSTM-based autism detection model for children using Federated Learning (FL). In the model, temporal and spatial information is extracted from the facial CNNs are highly adept at using convolutional filters to extract spatial features from images. LSTM networks are a specific type of Recurrent Neural Network (RNN) that is ideal for processing time-series or sequences because it can identify long-term relationships in sequential data. This architecture uses CNN layers to extract spatial information from important indications that are important for detecting ASD, like eye patterns, gestures, and facial expressions. After that, these features are sent to LSTM layers, which examine the time-dependent and sequential behavioral patterns associated with autism. Federated Learning allows the locally to train the model on its own dataset locally, sharing only model updates with a central server, thereby preserving data privacy while promoting diverse data contributions. According to experimental results using the proposed techniques, the federated CNN-LSTM model performs 4.3% better than the conventional centralized models because it has less overfitting and is more resilient to a range of data distributions. The model’s performance metrics further highlight its reliability, accuracy, precision, recall, and F1-Score values reaching 98.90%, 97.80%, 98.05%, and 98%, respectively, showing its potential for reliable ASD detection in children across diverse populations.