ThoraxAI
ThoraxAI is an advanced artificial intelligence-powered system designed for the automated detection and diagnosis of thoracic diseases using chest X-ray images. Developed with the goal of addressing the global shortage of radiologists and the rising demand for rapid and accurate diagnostics, ThoraxAI leverages deep learning using convolutional neural networks (CNNs) to analyze medical imaging data with high precision.
The system is trained on two large-scale, annotated chest X-ray datasets, namely NIH ChestX-ray14 and CheXpert, enabling it to recognize a wide spectrum of thoracic abnormalities, including pneumonia, tuberculosis, lung nodules, cardiomegaly, and pleural effusion. Through image preprocessing, feature extraction, and multi-label classification, ThoraxAI provides interpretable diagnostic outputs, including heatmaps and confidence scores, to assist clinicians in decision-making.
ThoraxAI is designed for integration into hospital radiology workflows, allowing real-time triage of urgent cases, reduction in diagnostic errors, and acceleration of reporting times. Its modular architecture supports both on-premise deployment in medical institutions and scalable cloud-based solutions for remote or underserved regions. Beyond diagnosis, ThoraxAI incorporates explainable AI components to enhance trust and usability among healthcare professionals by offering visual explanations of its predictions.
By combining technological innovation with practical clinical utility, ThoraxAI addresses key challenges in global healthcare, especially in low-resource settings by offering a scalable, efficient, and cost-effective diagnostic aid that empowers radiologists, reduces workload, and ultimately improves patient outcomes.
Deep Learning Models for Canine Dermatology: A Comparative Study Focused on Sri Lanka
Canine skin diseases, including canine scabies, fungal infections, and hypersensitivity allergies, are prevalent and require accurate diagnosis for effective treatment. However, traditional veterinary diagnostic methods face challenges such as similar clinical symptoms, limited accessibility to veterinary professionals, and misdiagnosis due to lack of expertise, particularly in rural areas of Sri Lanka. To address these challenges, the study explores deep learning-based classification for automated detection, addressing limitations in traditional veterinary diagnostics. A custom dataset of 2,511 images was collected and augmented to 12,565 images to improve model generalization. A comparative analysis was conducted using five deep learning models, including a baseline CNN, Hybrid CNN + ResNet50, ResNet152V2, EfficientNetB1, and MobileNetV3. The models were trained and evaluated based on their accuracy, loss, and generalization performance. The results indicate that EfficientNetB1 achieved the highest validation accuracy (98.96%) with a low loss (0.1278), followed closely by MobileNetV3 (98.76% accuracy). The Hybrid CNN + ResNet50 model balanced accuracy and efficiency, whereas the baseline CNN exhibited overfitting. Findings suggest EfficientNet-B1 and MobileNetV3 as optimal models for real-world deployment due to their accuracy and computational efficiency. Future work will focus on expanding the dataset, improving model interpretability using explainable AI techniques, and optimizing models for real-time deployment in veterinary applications. This research contributes to the advancement of AI-powered diagnostic tools that can assist veterinarians and pet owners in early and accurate detection of canine skin diseases.
QuanNetDetect – Quantum Hybrid Deep Learning Model Framework for Detecting Encrypted TLS Malicious Network Traffic
The research project is proposed as a novel approach for detecting encrypted malicious TLS network traffic by using Quantum Deep Learning. Without decrypting the traffic, by using the metadata (packet-level) information alone, the model can do the detection. The project involved multiclassification on TLS-based encrypted attacks and binary classification over TLS-based malicious network traffic by using Quantum Computing and Quantum Deep Learning.
MADQN-AV : A Multi-Agent Deep Reinforcement Learning framework for Emergent Cooperation and Conflict Resolution in Autonomous Vehicle Intersection Navigation
“The increasing adoption of autonomous vehicles (AVs) presents significant challenges in
managing intersections efficiently while minimizing conflicts. Traditional centralized traffic
control systems struggle with scalability and adaptability, whereas decentralized approaches
often face coordination and safety issues. Ensuring real-time decision-making and cooperation
among AVs in dynamic, multi-agent environments remains a critical challenge.
This research investigates a Multi-Agent Deep Reinforcement Learning (MADRL)
framework for Emergent Cooperation and Conflict Resolution in AVs. The approach utilizes
decentralized learning, allowing AVs to develop cooperative driving strategies based on
observed behaviors and shared environmental interactions, rather than explicit communication.
The proposed MADRL model is evaluated against a baseline algorithm, MADQN-AV, using key
performance metrics, including collision rate, throughput, average delay, and decision accuracy.
Experimental results indicate that the MADQN framework significantly reduces collision
rates, enhances traffic efficiency, and maintains real-time decision-making capabilities with
minimal computational overhead. Comparative analysis with MADQN-AV demonstrates
superior scalability and adaptability, validating the potential of decentralized multi-agent
learning in autonomous traffic systems.”
A Comparative Analysis of Classical and Deep Learning Models for Provider-Level Fraud in Healthcare
This Project compares several machine learning models to evaluate it’s performance on identifying fraudulent claims
ArgusEye: A Federated Learning Based Intrusion Detection System with a Focus on Privacy Preservation
“ArgusEye is a privacy-preserving Intrusion Detection System (IDS) designed to detect network anomalies and attacks without compromising user data. Traditional IDS solutions typically rely on centralized machine learning models that require the collection of raw data from client nodes. While effective, this centralized approach raises serious privacy concerns, increases the risk of data breaches, and potentially violates compliance requirements in sensitive domains such as healthcare and finance.
ArgusEye addresses these limitations through the integration of Federated Learning (FL) and Differential Privacy (DP). Federated Learning enables decentralized model training by allowing each client device to train on its local data, sharing only model updates, and not the data itself, with a central server. To enhance privacy further, Differential Privacy is applied by adding noise to the aggregated updates, making it difficult to infer any single user’s contribution.
The system was developed and evaluated using a widely used benchmark dataset and was designed to simulate multiple clients operating in a federated learning environment. Experiments were conducted to study the impact of differential privacy on model performance and training time. ArgusEye achieved strong detection metrics, which demonstrate that federated learning, when combined with privacy-preserving technologies, can offer a practical and effective solution for intrusion detection in privacy-sensitive environments.”
Predicting the efficiency of apparel production using ML model
“This project focuses on predicting the efficiency of apparel production lines using machine
learning, with the goal of classifying lines into Efficient, Low Efficient, or Inefficient
categories. Efficient production is critical in the apparel industry, yet identifying
underperforming lines in real time remains a challenge due to the complexity of operational
factors like SMV,lead times, and production hours.
The project adopts a data-driven approach using a historical production dataset. After
preprocessing and exploratory analysis, key features were selected and engineered to enhance model performance. PyCaret, a python based machine learning library, was used to train and compare several classification algorithms, including Logistic Regression, Random Forest, LightGBM, and XGBoost. Based on accuracy, F1-score, and AUC, the XGBoost classifier was chosen as the final model due to its superior performance and robustness.
Hyperparameter tuning was applied to optimize the model, which achieved a final accuracy of 79% and a macro F1-score of 0.84.To ensure generalizability, 10-fold stratified cross-validation was used, and the model was tested on unseen data.
A Flask web application was developed to make the model accessible to non-technical users. This allows production managers to input planning data and receive immediate efficiency classifications, supporting proactive decision-making.
Overall, the project demonstrates how machine learning can enhance operational efficiency in apparel manufacturing. Though effective, the system could be improved by incorporating real- time ERP integration and expanding the dataset for broader applicability.”
BlurFix : Efficient Image Blind Motion Deblurring Using Generative Adversarial Network
Most real-world images are degraded by motion blur due to either camera shake or fast moving objects. This prevails as a great challenge for many applications in computer vision such as object detection and scene understanding, which really require clear images. Blind motion deblurring aims to restore sharp images from blurry ones without any prior knowledge of the blur kernel. Real and effective deblurring remains a big challenge due to detail retention and avoiding artifacts. Even though the recent advances in deep learning and, particularly, Generative Adversarial Networks have shown some promising work for this task, they utilize a high computational cost. This work presents a Mobile Vision Transformer (MobileViT) enhanced with gradient Wasserstein GAN penalty (WGAN-GP) efficient blind motion deblurring framework pairing a spectrally normalised discriminator for stable training with a U-Net style generator with MobileViT blocks for multi-scale feature fusion and global-local representation learning. The proposed architecture combines important developments including a hybrid CNN-Transformer generator using inverted residual blocks and skip connections to preserve high-frequency details, a perceptual loss formulation including VGG-16 features alongside adversarial and L1 losses (Mean Absolute error) and lightweight patch-based discrimination using spectral normalising for enhanced training stability. Using the GOPRO dataset, experiments show that our model maintains a lower computational complexity than equivalent GAN-based deblurring techniques while obtaining competitive PSNR/SSIM measurements.
mitoMatch – A Machine Learning Approach to Identify Human Relatedness Using Mitochondrial DNA Hypervariable Region I and II
This project presents an interdisciplinary approach that combines genomics and computing to identify human relatedness by predicting an individuals ethnicity and geographic region using mitochondrial DNA (mtDNA) Hypervariable region 1 and 2. Unlike nuclear DNA, mtDNA is maternally inherited and takes a long period of time to degrade. The ethnicity model is a Gradient Boosting model with accuracy 95% and the geographic location is a Random Forest model with accuracy 90%. The data obtained from GenBank to train the model has also been validated to prove that there is variation between the samples using Analysis of Molecular Variance (AMOVA) and a web application has been developed using React and Flask api to integrate the machine learning model.
WristGuard: A Deep Learning Approach for Detection and Classification of Wrist Fractures in Athletes
“The wristGuard application employs a hybrid approach, combining deep learning models for
feature extraction and stacking ensemble method for classification of wrist fracture types.This
algorithm combines deep learning-based feature extraction with a stacking ensemble classification
approach to improve wrist fracture detection. It utilizes four pre-trained CNNs, namely MobileNet,
ResNet50, DenseNet121, and InceptionV3, to extract high-level features from grayscale X-ray
images, which are then used by XGBoost classifiers for fracture classification. “