Smart Computing and Systems Engineering - 2025 (SCSE 2025)

Permanent URI for this collectionhttp://repository.kln.ac.lk/handle/123456789/30037

Browse

Now showing 1 - 20 of 32

Enhancing Sign Language Communication: Advanced Gesture Recognition Models for Indian Sign Language
(Department of Industrial Management, Faculty of Science, University of Kelaniya., 2025) Kumar, S. G. S.; Abbass, J.
Globally, millions of individuals experience varying degrees of hearing impairment, creating an urgent demand for effective communication solutions. The limited number of proficient sign language users exacerbates this challenge. Recent advancements in machine learning provide promising avenues to address this issue. This study introduces an innovative automated system that translates one of the most popular sign languages, namely, Indian Sign Language (ISL), into English text using a webcam. Our comprehensive dataset includes ~1M images across 36 categories, covering digits (0-9) and alphabet letters (A-Z). The dataset features diverse gestures captured from various angles and performed by 6 individuals with different characteristics followed by data augmentation. We evaluated the effectiveness of 5 models, 3 standard and 2 customized respectively: (1) MobileNetV2, a pre-trained convolutional neural network (CNN) optimized for mobile applications, (2) VGG16, a well-established pre-trained deep learning model, (3) the standard CNN, (4) a custom-designed CNN tailored for ISL recognition, trained on 32x32 images for 20 epochs, and (5) Customized MobileNetV2 for ISL recognition retrained on 128x128 images for 20 epochs. Both customized models achieved an F1-Score of 94 whilst standard ones achieved an F1-Score of no more than 85. The comprehensive comparison underscores the enhanced accuracy and efficiency of our custom models, establishing them as a significant advancement in sign language recognition.
Predictive policing with neural networks: A big data approach to crime forecasting in Sri Lanka
(Department of Industrial Management, Faculty of Science, University of Kelaniya., 2025) Nauzad, H.; Dayawansa, D.; Dias, N.; Haddela, P. S.; Ratnayake, S.
The surge in crime rates, particularly in urban regions, has underscored the importance of predictive policing within law enforcement strategies. This research introduces a neural network-based crime prediction model, specifically tailored to address the complexities of Sri Lanka’s crime landscape. By combining big data analytics with advanced machine learning methods—including ensemble models such as Random Forest and Gradient Boosting, alongside Artificial Neural Networks (ANNs)—our study presents a robust framework to forecast crime incidents, locations, and time spans. While neural networks excel in predictive accuracy, their “black-box” nature can hinder practical applications in critical fields like law enforcement. To address this, our model integrates Explainable AI (XAI), making the decision-making process of the system transparent and interpretable for end-users. XAI helps break down complex neural network predictions, ensuring trust and clarity in the model’s insights. With a prediction accuracy rate of 85%, this approach demonstrates substantial potential to improve crime prevention efforts and optimize resource allocation. Our research not only highlights the predictive strengths of neural networks but also showcases the essential role of interpretability for deploying these models effectively in real-world policing.
Hybrid CNN-LSTM Framework for Robust Speech Emotion Recognition
(Department of Industrial Management, Faculty of Science, University of Kelaniya., 2025) Shaik, A.; Reddy, G. P.; Vidya, R.; Varsha, J.; Jayasree, G.; Sriveni, L.
A key component of emotional computing is voice Emotion Recognition (SER), which is concerned with recognizing and categorizing human emotions from voice data. Human communication is heavily influenced by emotions, and giving machines the ability to recognize these emotional states improves their capacity for intelligent and sympathetic interaction. A reliable SER system that accurately classifies emotions using a hybrid Convolutional Neural Network (CNN) and Long Short-Term Memory (LSTM) model is presented in this research. The suggested approach integrates sophisticated feature extraction methods that efficiently capture both temporal and spatial emotional patterns in speech, including Mel-Frequency Cepstral Coefficients (MFCC), pitch, and chroma data. Metrics including accuracy, precision, recall, and F1-score were used to assess the system on two common datasets, RAVDESS and EMO-DB. The trial findings show that the hybrid CNN-LSTM model outperformed traditional machine learning approaches, with an overall accuracy of 89.4%. The system also demonstrated resistance to external noise and emotional overlap, making it appropriate for real-world applications.
Enhancing network intrusion detection with stacked deep and reinforcement learning models
(Department of Industrial Management, Faculty of Science, University of Kelaniya., 2025) Kalpani, N.; Rodrigo, N.; Seneviratne, D.; Ariyadasa, S.; Senanayake, J.
This study investigates the effectiveness of Ensemble Learning (EL) techniques by integrating reproducible Deep Learning (DL) and Reinforcement Learning (RL) models to enhance network intrusion detection. Through a systematic review of the literature, the most effective DL and RL models from 2020 to 2024 were identified based on their F1 scores and reproducibility, focusing on recent advancements in network intrusion detection. A structured normalisation and evaluation process allowed for an objective comparison of model performances. The best performing DL and RL models were subsequently integrated using a stacking ensemble technique, chosen for its ability to combine the complementary strengths of the DL and RL models. Experimental validation in a benchmark dataset confirmed the high accuracy and robust detection capabilities of the model, outperforming the individual DL and RL models to detect network intrusions in multiple classes. This research demonstrates the potential of ensemble methods for advancing Intrusion Detection Systems (IDSs), offering a scalable and effective solution for dynamic cybersecurity environments.
A Hybrid Architecture with Efficient Fine Tuning for Abstractive Patent Document Summarization
(Department of Industrial Management, Faculty of Science, University of Kelaniya., 2025) Jayatilleke, N.; Weerasinghe, R.
Automatic patent summarization approaches that help in the patent analysis and comprehension procedure are in high demand due to the colossal growth of innovations. The development of natural language processing (NLP), text mining, and deep learning has notably amplified the efficacy of text summarization models for abundant types of documents. Summarizing patent text remains a pertinent challenge due to the labyrinthine writing style of these documents, which includes technical and legal intricacies. Additionally, these patent document contents are considerably lengthier than archetypal documents, which complicates the process of extracting pertinent information for summarization. Embodying extractive and abstractive text summarization methodologies into a hybrid framework, this study proposes a system for efficiently creating abstractive summaries of patent records. The procedure involves leveraging the LexRank graph-based algorithm to retrieve the important sentences from input parent texts, then utilizing a Bidirectional Auto-Regressive Transformer (BART) model that has been fine-tuned using Low-Ranking Adaptation (LoRA) for producing text summaries. This is accompanied by methodical testing and evaluation strategies. Furthermore, the author employed certain meta-learning techniques to achieve Domain Generalization (DG) of the abstractive component across multiple patent fields.
Bridging Linguistic Gaps: A Review of AI- Driven Speech-to-Speech Translation for Sinhala and Tamil in Sri Lanka
(Department of Industrial Management, Faculty of Science, University of Kelaniya., 2025) Dilshani, I.; Chandrasena, M.
Sri Lanka maintains two official languages which constitute Sinhalese and Tamil making it linguistically diverse. Almost all communication requires effective dialogue between Sinhalese-speaking and Tamil-speaking communities especially when operating through real-time speech-to-speech translation. The current version of Speech-to-Speech Translation (S2S) solutions serves a useful purpose yet faces three major limitations, including internet dependence, performance difficulties in loud environments, and unnatural Text-to-Speech (TTS) outputs. The Automatic Speech Recognition (ASR) systems from CallTran along with Android-based solutions through Google APIs and PocketSphinx, struggle with flexible operations when processing different accent varieties. Furthermore, the Machine Translation (MT) system performs poorly in achieving semantic relevance due to the scarcity of parallel corpora. The combination of ASR, MT and TTS systems produces performance delays and misinterpretation issues, which interfere with real-time functionality. This review examines current models, highlights theoretical and practical gaps, and proposes directions for future research, followed by a comparison of existing approaches. The research requires attention to three essential gaps, including bilingual dataset annotation tasks alongside offline functionality and natural voice synthesis development. We propose future research directions to establish massive bilingual datasets as well as implement noise-resistant ASR models using self-supervised approaches such as Whisper and Wave-to-Vector v2 (Wav2Vec2) and the fine-tuning of multilingual MT models like Multilingual Bidirectional and Auto-Regressive Transformer (mBART) for Low Resource Sinhala-Tamil Language translation systems. Additionally, TTS models like Tacotron, FastSpeech and Coqui TTS should be optimized for prosody and intonation.
Virtually restoring headless Buddha statues using Deep Learning Techniques
(Department of Industrial Management, Faculty of Science, University of Kelaniya., 2025) Liyanage, M. M.; Induka, L. P. D. P.; Dissanayake, P. P.; Wijeywickrama, W. K. D. A.; Kulathilake, K. A. S. H.
Buddha statues are undoubtedly important for culture and history worldwide, particularly in Sri Lanka, where Buddhism is the major religion and represents the country’s rich historical heritage. Many historical Buddha statues from the earliest eras, such as those in Anuradhapura, have been destroyed or degraded due to natural disasters, vandalism, or aging. It is common for body parts, especially the head, to be lost or damaged in these statues. While various physical restoration methods exist, they are often expensive and may reduce the historical or artistic value of the statues. Virtual restoration methods, on the other hand, offer a non-invasive and affordable solution for preserving and reconstructing these heritage assets. This study presents a specially designed deep learning model using a U-Net architecture to segment the body of seated headless Buddha statues by removing the background from input images, trained on a custom dataset of Buddha statues found in Anuradhapura, Sri Lanka. These segmented images eliminate unnecessary details and provide a clean statue body for subsequent restoration tasks. The head inpainting phase, based on the segmented image, will be carried out using a GAN-based architecture, an extension of Pix2Pix, which virtually inpaints a matching head for the given statues. This two-stage approach—segmentation followed by inpainting—provides a complete pipeline for the virtual restoration of headless Buddha statues, maintaining cultural integrity while minimizing physical intervention. The proposed methodology demonstrates efficiency and serves as a valuable tool in heritage conservation and the virtual reconstruction of culturally significant artifacts.
Predicting Medical Drug Sales in a Specific Area for Categorical Drugs using Time Series Forecasting
(Department of Industrial Management, Faculty of Science, University of Kelaniya., 2025) Ekanayake, S. B.; Lakmal, G.; Perera, A.; Nasmeen, M.; Vimanshani, P.; Chandrasena, M.
Accurately forecasting pharmaceutical drug sales is a significant challenge faced by many firms, particularly in Sri Lanka, where factors such as seasonality, weather, local health crises, importation issues, currency fluctuations, and economic instability affect inventory management. These challenges often lead to frequent conditions of either shortages or overstocking of drugs, which adversely affect healthcare delivery and business profitability. This study addresses this issue through the development of a data-driven system using machine learning to predict drug sales efficiently and accurately. This work involved gathering sales data from local pharmacies, performing pre-processing steps, and implementing a time-series forecast using the SARIMA model, which works efficiently with seasonal variations in sales data. A locally hosted, user-friendly web application was developed using the Flask framework to present these predictions in a readable format for pharmacists and drug sellers. The system was also validated on an external dataset, demonstrating high accuracy in the forecasted sales, which helped improve inventory management practices. The proposed system reduces drug shortages, minimizes wastage due to expiration, and enhances supply chain efficiency, thereby improving healthcare delivery and business outcomes. This research provides evidence of the opportunity to leverage pharmaceutical sales data to identify disease trends and inform public health strategies. The model can be further improved and applied in various aspects by including additional variables. This research bridges gaps in supply chain management, improving the availability of medications and making inventory management more predictable, benefiting both public health and industry stakeholders.
Hybrid Model-Based Automated Exterior Vehicle Damage Assessment and Severity Estimation for Insurance Operations
(Department of Industrial Management, Faculty of Science, University of Kelaniya., 2025) Jayagoda, N. M.; Kasthurirathna, D.
After a vehicle accident, insurance companies face the critical task of assessing the damage sustained by the involved vehicles, a process essential for maintaining the insurer's credibility, building consumer trust, and meeting legal and ethical obligations. This assessment is crucial for ensuring clients' financial protection and proper compensation, upholding the integrity of the insurance process. Traditionally, evaluations have been conducted through manual inspections by experienced professionals who meticulously document vehicle damage. Despite its thoroughness, this approach suffers from significant inefficiencies, high costs, and extended time requirements. Moreover, the method is vulnerable to human errors and subjective biases, which can result in inflated valuations. To overcome these challenges, this research introduces an innovative system designed to leverage technology for analyzing images of damaged vehicles uploaded by the user. This system aims to accurately identify the damaged external components, assess the severity of the damage, and determine the repair needs based on the compromised sections of the vehicle. The findings reveal that the hybrid model used in this research is capable of determining vehicle damage severity with an overall accuracy of 73.3%. This level of accuracy demonstrates the model's robust capability to effectively navigate and analyze complex damage patterns, underscoring its practical applications. By accurately determining damage levels on the first assessment, the model reduces the need for further assessments and disagreements, which frequently cause claim delays. This enhancement increases productivity, reduces administrative costs, and improves the customer experience, resulting in a more efficient, transparent, and satisfactory resolution of vehicle insurance claims.
Detection of Cyberbullying to Reduce Mental Health Problems using Machine Learning Algorithms
(Department of Industrial Management, Faculty of Science, University of Kelaniya., 2025) Perera, M. V. V.; Piyumal, K. M.
Social networks and other online platform services are where people are more likely to experience issues with cyberbullying, including kids, young and older adults who are addicted to them. Cyberbullying is an activity that takes place on digital platforms where victims are threatened or bullied individually or in groups by messages or comments online. Various cyberbullying detection techniques are continuously used on social media platforms; however, not all online platform services follow those mechanisms, which may lead to psychological problems that can cause depression and even suicide because people are unaware of taking action to prevent it. Many past cyberbullying detection studies used small datasets and omitted disclosing the total number of features used to train the model. To fill this gap, this study explores how model performance changes with the feature count and what happens when the dataset size increases. Therefore, two cyberbullying datasets with a combined total of 47,183 and 120,556 records were used, containing suspicious activities on Twitter and Facebook that most commonly belong to the cyberbullying category. To compare performance metrics of each model, three methods for feature extraction and three classifiers were used, namely Logistic Regression (LR), Random Forest (RF), and Support Vector Machine (SVM). The highest accuracy for the models created utilizing 47,183 data under the three feature extraction approaches was 94.43%, while the highest accuracy for the 120,556 data was 89.96%.
Impact of Rainfall Patterns on Flood Dynamics in the Gin River Basin, Sri Lanka: A Data Warehouse Approach
(Department of Industrial Management, Faculty of Science, University of Kelaniya., 2025) Rathnayake, R. M. K.; Ratnayake, P. S. S.; Dilanka, W. A. A.; Madhusanka, K. P. P.; Peiris, H. S. S.; Herath, H. M. M. V.; Asanka, P. P. G. D.
This research study investigates the impact of rainfall patterns on flood dynamics in the Gin River basin, located in the Southwestern area of Sri Lanka, with a multidimensional data warehouse application to integrate and analyze hydro-climatological data across various temporal scales. Granular exploration of rainfall patterns and water levels is enabled through Online Analytical Processing (OLAP) architecture and Key Performance Indicators (KPI). The findings of the study show that the Southwest (SW) monsoon is the primary driver in flood occurrences in the basin, with the highest mean rainfall (80.94 mm/day) and statistically significant at p < 0.05 and p < 0.01 compared to other seasons. Transitional seasons, particularly the 2nd Inter Monsoon (IM2), also contributed to floods. Generalized Extreme Value (GEV) analysis revealed increasing extreme rainfall risks, especially in SW and IM2. Moreover, the land use changes such as deforestation and rapid urbanization in the basin have caused a growth of flood risks by increasing the surface runoff. This approach offers a scalable data warehouse applicable to any other flood-prone area, assisting effective flood risk management.
Deepfake Detection Using a Hybrid Deep Learning Approach with Swin Transformers and ConvNeXt
(Department of Industrial Management, Faculty of Science, University of Kelaniya., 2025) Amarasinghe, H. M. S.; Kasthuri Arachchi, S. P.
The rapid advancement of artificial intelligence has led to the proliferation of deep-fake technology, which poses significant challenges to digital security and trust. As deepfakes become increasingly sophisticated, there is an urgent need for effective detection methods that can accurately identify manipulated media across various platforms and contexts. Deepfakes are outcomes of advanced Artificial Intelligence algorithms such as Generative Adversarial Networks, and have become a danger to digital integrity, personal privacy, and public trust. With continuous advancements in the techniques for generating deepfakes, conventional methods of detection based on the artifact analysis of visuals and inconsistencies in physiological signals have started to wear out. This paper presents a hybrid deep learning model that works between Swin Transformers and ConvNeXt architectures for better detection. The proposed model achieves better detection accuracy and robustness by leveraging Swin Transformers' hierarchical feature extraction capabilities and the efficient processing strengths of ConvNeXt. Obtained results on the "Deepfake and Real Images" dataset demonstrated performance of 95.6% accuracy, 97.4% precision, 93.7% recall, and a 95.2% F1 score. The hybrid model is more effective than existing ones, demonstrating their potential for concrete application in social media platforms, news agencies, and digital forensics to help fight misinformation and preserve digital trust.
Accelerating Meta-Learning with the Enhanced Reptile Algorithm for Rapid Adaptation in Neural Networks
(Department of Industrial Management, Faculty of Science, University of Kelaniya., 2025) Shree Smeka J.; SheejaKumari V.; Vijaya Raj M.; Santhosh Kumar S. P.; Angalaeswari S.
Reptile is an innovative meta-learning approach that improves the process of neural networks training across diverse tasks. Reptile differs from the preference gradient strategies. It searches for a model weight vector which can enable a model to learn a new set of tasks with a small number of weight updates. This is accomplished via a first-order optimization process which makes it less intricate than other strategies like model-agnostic meta-learning MAML. Reptile is said to sample a number of tasks and then trains the model on each of the tasks via performing a series of a few gradient steps, systematically updating the model towards the average gradient direction across all the tasks. This enables the model to generalize well with new tasks trained with few iterations, thus proving beneficial in few-shot learning. Reptile enables faster adaptation of the model by focusing on the learning of better initial parameters thereby lowering computational overhead and the training duration. The algorithm is noted for its ability not only to learn but also to adapt itself to new tasks in a short period of time, which greatly extends the scope of its application, especially in areas where data is scarce. Examples of these areas are robotics, personalized advertisements, and decision-making approaches that need to operate in real time. Reptile Algorithm builds good on gradient-based approaches and can spearhead volumetric applications of meta reasoning by being exceptionally efficient, scalable and less computationally intensive.
A Unified Pipeline for Improving Financial Aid Eligibility Predictions in Deep Learning Models
(Department of Industrial Management, Faculty of Science, University of Kelaniya., 2025) Kishanthan, S.; Hevapathige, A.
This paper systematically evaluates the performance of widely used deep learning architectures for financial aid prediction, identifying critical bottlenecks that hinder optimal performance. To address these limitations, we further propose a novel pipeline that enhances deep learning models by incorporating three key components: text vectorization, data equalization, and adaptive feature recalibration. This pipeline improves the models’ representational power and predictive accuracy, offering seamless integration with existing architectures. It significantly boosts performance in predicting financial aid eligibility, providing up to a 145% increase in balanced accuracy and a 45% increase in F1 score.
Gait pattern analysis for a weight carrying hexapod ant robot using reinforcement learning
(Department of Industrial Management, Faculty of Science, University of Kelaniya., 2025) Karunarathna, J. A. T. D. B.; Mohamed Saki, W.; Kanesalingam, S.; Prasanga, D. K.; Weerasinghe, W. A. B. G. H. B. P.; Abeykoon, A. M. H. S.; Ruwanthika, R. M. M.
Robot hexapods are coming into focus in robotics fields due to their stable base and versatility in handling different terrains. The current study aims at identifying gait patterns to enhance efficiency and stability under dynamic payload conditions. It investigated six payload conditions: no payload, central payload, and four asymmetric payloads, which were front-left and back-left and front-right and back-right. To enable the dynamic modification of gait patterns, Proximal Policy Optimization (PPO), which is a type of reinforcement learning, was used in order to foster efficient and stable forward propulsion. To train and simulate a robot, Brax, an open-source physics simulation environment, was used under different payload conditions. It was shown that gait adaptation to loading distribution is achievable, while bilateral loading causes energy expenditures to grow. The research on hexapod movement is useful in the advancement of the field of bio-inspired robotics; it provides ideas for increasing hexapod mobility with unequal weight loadings and also helps to further extend hexapod robots’ usability.
Sector-Specific Electricity Demand Forecasting in Sri Lanka Using Deep Learning and Hybrid Models
(Department of Industrial Management, Faculty of Science, University of Kelaniya., 2025) Fernando, G.; Hewapathirana, I.
Electricity demand forecasting is important in addressing the growing challenges in the energy sector. Traditional forecasting models often fail to capture the complex interplay of external factors, particularly in countries like Sri Lanka, where electricity demand in different sectors, such as the domestic and industrial sectors, which collectively account for over 65% of total electricity consumption, exhibit unique patterns. This study introduces a sector-specific approach to electricity demand forecasting in Sri Lanka's domestic and industrial sectors by employing Deep Learning models and integrating socioeconomic and weather variables to enhance accuracy. Key predictors for each sector were identified using Random Forest-based feature selection. A multivariate multi-step Long Short-Term Memory (LSTM) model, the Seasonal Autoregressive Integrated Moving Average (SARIMA) model, and a hybrid SARIMA–LSTM model were implemented to assess short-term and long-term forecasting capabilities. Results demonstrated that the 2-step LSTM model and the hybrid SARIMA–LSTM model outperformed the SARIMA model. Integrating diverse datasets enhances forecasting accuracy, providing actionable insights for sustainable energy planning and resource optimization. The study’s originality lies in its sector-specific focus, incorporating weather and socioeconomic features and the innovative use of hybrid models to address the unique electricity demand patterns in Sri Lanka.
Enhancing Network Security using MachineLearning for Automated Anomaly-based Intrusion Detection Systems for IoT Environment
(Department of Industrial Management, Faculty of Science, University of Kelaniya., 2025) Mahamud, N.; Uddin, M. J.; Sumaiya, U.
The rapid expansion of the Internet of Things (IoT) has revolutionized modern life, offering unparalleled automation and seamless interconnectivity between devices, often operating without user intervention. However, this convenience comes with a significant trade-off: increased susceptibility of IoT devices to cyberattacks, which can result in severe consequences if not promptly addressed. To tackle this pressing challenge, our study proposes innovative strategies powered by machine learning algorithms, achieving an exceptional 99.97% detection accuracy and a 0.0% false positive rate. Leveraging the Bot-IoT dataset for evaluation, our approach demonstrates marked improvements over existing detection methodologies. Furthermore, its adaptability to diverse IoT applications underscores its potential as a transformative advancement in IoT security.
Integrating CNN-GRU-BiLSTM for Robust Schizophrenia Detection using Deep Learning
(Department of Industrial Management, Faculty of Science, University of Kelaniya., 2025) Reetha, L.; Gnanajeyaraman, R.
Early detection of schizophrenia is critical but challenging due to its complex presentation. In order to interpret EEG data, this work proposes a hybrid deep learning model that combines Convolutional Neural Networks (CNNs), Gated Recurrent Units (GRUs), and Bidirectional Long Short-Term Memory (BiLSTM) networks with an attention mechanism. The model captures spatial, temporal, and contextual dependencies, achieving a classification accuracy of 99.30%, outperforming existing methods like CNN-LSTM and CNN-BiLSTM. By integrating spatial feature extraction, temporal dynamics, and attention for interpretability, the model offers robust, efficient, and transparent diagnostics. Verified on TUH EEG Corpus and CHB-MIT EEG datasets, it demonstrates the potential of deep learning models based on EEG for accurate and scalable early schizophrenia identification, opening the door for revolutionary uses in mental health diagnostics.
Deep Learning based Screen Display Fault Detection System for Vehicle Infotainment Applications
(Department of Industrial Management, Faculty of Science, University of Kelaniya., 2025) Ramesh, B.; Dheeba, J.; Raja Singh, R.
Modern vehicles are integrated with in-vehicle infotainment systems and are subject to software faults. This paper explores the application of deep learning algorithms to identify visual defects in infotainment systems and automatically document the issues. A real-time capable framework is deployed, delivering immediate feedback on detected defects. The proposed system performs thorough analysis, automatically summarizes detected defects, and generates detailed reports, significantly reducing manual documentation effort and supporting faster decision-making. The performance of the developed models is evaluated using Convolutional Neural Networks (CNN) and Artificial Neural Network (ANN) classifiers. Experimental results demonstrate the superior performance of the CNN model, achieving a training accuracy of 82.21% with an F1 score of 0.85, and a testing accuracy of 80.51% with an F1 score of 0.811. In comparison, the ANN model achieves a training accuracy of 70.18% with an F1 score of 0.7314, and a testing accuracy of 69.32% with an F1 score of 0.705.
Federated SCFGWO for Secure and High-Accuracy Brain Tumor Detection Across Multi-Center MRI
(Department of Industrial Management, Faculty of Science, University of Kelaniya., 2025) Alphonsa, J.; Kumari, V. S.
Privacy preservation and security concerns in medical imaging have led to the implementation of federated learning (FL) as an alternative to centralized machine learning approaches. This study introduces a novel Sine and Cosine Fitness Grey Wolf Optimization (SCFGWO) algorithm integrated into an FL framework to enhance the accuracy and efficiency of brain tumor detection across multi-center MRI datasets. SCFGWO addresses traditional Grey Wolf Optimization (GWO) limitations, such as early convergence and local optima, by incorporating adaptive sine–cosine strategies for improved global exploration and local exploitation. Compared to existing methods, the proposed SCFGWO-based framework achieves superior accuracy (97.2%), Dice Similarity Coefficient (0.94), and Intersection over Union (0.90) while maintaining computational efficiency (0.015s per slice). Innovative privacy-preserving methods, including differential privacy and homomorphic encryption, ensure data security across the dataset. These contributions establish SCFGWO as a robust optimization tool for federated learning applications in medical imaging, offering improved performance and enhanced data security.