http://dlib.iit.ac.lk/xmlui/handle/123456789/2885 2026-04-28T14:04:37Z http://dlib.iit.ac.lk/xmlui/handle/123456789/2949 Self-Supervised Learning for Automated Fracture Detection in Radiographic Images Perera, Lahiru Bone fractures are a critical medical condition requiring early and accurate detection to ensure timely treatment, yet conventional analysis relies on manual interpretation, which is time- consuming and prone to error (Sharma et al., 2025). While deep learning models have been applied to this problem, they often struggle with the small and complex datasets typical in medical imaging, leading to unreliable results (Alwzwazy et al., 2025). The primary limitations hindering their clinical adoption are a heavy dependence on large, expert-annotated datasets, which are difficult to acquire, and a lack of model transparency, which erodes clinical trust (Alwzwazy et al., 2025). To address this gap, this research designed and developed a novel, end-to-end fracture detection system built upon a Vision Transformer (ViT) architecture. This approach diverges from traditional supervised methods by leveraging a domain-specific Self-Supervised Learning (SSL) strategy, a technique that has shown significant potential to enhance clinical diagnostics (Wang and Siddiqui, 2024). The core of the solution is a two-stage process. First, a standard ViT- Base/16 encoder was pre-trained on a large corpus of unlabeled musculoskeletal radiographs from the MURA dataset using a Masked Autoencoder (MAE) framework. This forces the model to learn rich, high-level semantic features of radiographic anatomy without requiring any human-provided labels. Subsequently, this pre-trained encoder was fine-tuned for fracture classification using a smaller, labeled subset of the MURA dataset. The developed SSL-ViT model was systematically evaluated on a hold-out validation set, demonstrating the viability and effectiveness of the proposed approach. The system achieved a validation accuracy of 86.90% and an Area Under the Curve (AUC-ROC) score of 0.8686, indicating a strong capability to distinguish between fractured and non-fractured cases. Analysis of the training dynamics confirmed that the SSL pre-training provided a robust foundation, enabling the model to learn effectively from limited labeled data. These results validate that the combination of domain-adaptive self-supervised learning with Vision Transformers presents a promising pathway toward creating more data-efficient, accurate, and trustworthy AI tools for clinical diagnostics. 2025-01-01T00:00:00Z http://dlib.iit.ac.lk/xmlui/handle/123456789/2948 A Hybrid Approach to Fake News Detection Using Explainable AI and Multimodal Content Analysis Elpitiya Badalge, Tharindi The rapid spread of fake news across digital platforms has occurred as a significant risk to societal trust, community safety, and political stability. Misinformation campaigns often exploit various media formats such as text, images, and videos making it increasingly problematic to distinguish between trustworthy and misleading news content. Traditional approaches of false news detection primarily focus on textual analysis, leaving gaps in detecting multimodal misinformation The proposed system has used advanced multimodal models CLIP (OpenAI), ViLT, BERT, ResNet50, and VisualBERT to evaluate. Text is processed using BERT, while ResNet50 and SAFE handle image features. ViLT and VisualBERT model text-image relationships, and CLIP aligns visual and textual semantics. After evaluating all models, the best-performing ones are combined to improve accuracy and generalization. Explainability is ensured through SHAP or LIME, helping users understand the reasoning behind each prediction. This approach is demonstrated through a prototype, assessed using standard metrics like accuracy, recall, precision, and F1-score. The initial model, without zero-shot learning, achieved strong performance on the Twitter dataset with an accuracy of 97.61%, precision of 97.62%, recall of 97.61%, and F1-score of 97.61%. When evaluated with zero-shot learning on the FakeNewsNet dataset, the enhanced model achieved 93.91% accuracy. The proposed solution promises to be an effective tool in combating misinformation by providing a robust, explainable system for fake news detection across diverse media formats. Future work includes expanding the model’s capabilities to handle real-time data and multimedia content, further improving the model's efficiency and adaptability in dynamic environments. 2025-01-01T00:00:00Z http://dlib.iit.ac.lk/xmlui/handle/123456789/2947 NimbusGuard: A Novel Intelligent Orchestration Framework for Proactive Recovery and Performance Optimization in Kubernetes Wanigasooriya, Chamath Reactive, threshold-based autoscaling mechanisms in Kubernetes are inadequate for modern cloud-native applications, as they fail to efficiently manage dynamic workloads, leading to performance degradation and inefficient resource utilization. This inherent "scaling lag" creates a critical need for an automated, intelligent, and proactive management paradigm. This research addresses this challenge by introducing NimbusGuard, a novel, intelligent orchestration framework designed for proactive recovery and performance optimization in Kubernetes. The proposed solution overcomes the limitations of traditional autoscalers by synergistically integrating a Long Short-Term Memory (LSTM) network for predictive forecasting with a Deep Q-Network (DQN) agent for adaptive, multi-objective decisionmaking. A key contribution of this work is the practical implementation of this intelligence within a stateful LangGraph workflow, which includes a crucial MCP safety validation layer to ensure all scaling decisions are orchestrated and validated, preventing system instability. In empirical benchmarks against industry-standard autoscalers, NimbusGuard demonstrated vastly superior performance, reducing the average time to scale by 80% compared to the default Horizontal Pod Autoscaler (HPA) and by over 33% compared to KEDA. This research contributes a production-aware framework that successfully bridges the gap between theoretical AI models and their practical, safe deployment, offering a tangible solution that enhances the performance, efficiency, and resilience of modern cloud-native applications. 2025-01-01T00:00:00Z http://dlib.iit.ac.lk/xmlui/handle/123456789/2946 Enhancing Communication for the Hearing-Impaired: A Vision-Based System for Translating SSL with Emotional Expression Segar, Arosh Communication barriers for the hearing-impaired community in Sri Lanka persist due to the lack of accessible, real-time Sri Lankan Sign Language (SSL) translation systems that capture emotional expressiveness. Existing systems mainly recognize hand gestures but overlook facial expressions, resulting in translations that miss emotional nuance and naturalness, limiting their effectiveness in daily and critical communication. This project presents a multimodal, emotion-aware SSL translation system combining Timesformer-based gesture recognition with MediaPipe facial and body landmark fusion. A facial emotion detection module using DeepFace extracts dominant emotions, which are converted into expressive speech via a Typecast API-powered emotional TTS. The backend is built as FastAPI microservices deployed on cloud platforms, integrated with a Flutter mobile app interface. The system employs deep learning, multimodal fusion, and user-centered design, validated through extensive training and mixed-method evaluations. The prototype achieved around 85% gesture recognition accuracy on SSL datasets, over 80% accuracy in emotion detection, and generated natural, context-aware speech with latency suitable for near real-time use. Balanced metrics like F1 score and precision-recall, alongside expert and user feedback, demonstrate the system’s robustness and improved communication clarity in both every day and emergency scenarios. 2025-01-01T00:00:00Z