Connected vehicle systems use sensors and machine learning to monitor road conditions in real time. They apply federated learning to preserve privacy while detecting road anomalies. Advanced models improve accuracy and scalability. The system enables efficient, privacy-preserving monitoring without centralized data sharing.
HandOsteoNet is a deep learning model for automated bone age assessment from hand X-rays. It improves accuracy using advanced feature extraction and attention mechanisms. The model performs real-time predictions with strong generalization across datasets. It also provides interpretable outputs by highlighting key anatomical regions.
MTPINN is a physics-informed deep learning model for classifying gas–liquid flow regimes. It integrates physical constraints with signal-based features to improve accuracy and reliability. The approach achieves strong performance while maintaining interpretability. It also enables real-time flow analysis and decision support.
A VLM-RAG pipeline is developed to classify sacroiliac joint STIR MRI for sacroiliitis detection. It integrates a vision-language model with guideline-based retrieval to improve interpretability and reliability. The system works without fine-tuning and achieves strong diagnostic performance. It provides explainable, clinically grounded predictions to support early diagnosis.
HierTAC-Net is a deep learning model for classifying gastrointestinal endoscopic images to support early colorectal cancer detection. It combines hierarchical attention and advanced feature extraction for high accuracy and robustness. The model performs real-time analysis and generalizes well across datasets and conditions. It also shows strong clinical performance compared to experts.
Adaptive BitFit is a lightweight fine-tuning method for histopathological image classification in cancer diagnosis. It updates only key bias parameters to reduce computational cost while preserving accuracy. The method performs well across multiple datasets with strong generalization. It also improves efficiency without significantly affecting diagnostic performance.
We developed ReliefNet, an integrated AI and blockchain-based humanitarian platform for real-time disaster severity assessment, automated reporting, donation forecasting, and transparent donation tracking. It combines multimodal AI models, XAI, a web-based disaster classification app, and a blockchain-secured donation management prototype.
AI-based contactless human weight estimation prototype that uses camera vision to estimate body weight from front and side-view images. The system combines computer vision, machine learning, and local embedded processing to provide a low-cost, privacy-preserving alternative to conventional weighing scales.
In this project, we developed AirSense-R, a mobile robot for indoor air quality mapping. Our robot moves around the room, collects air-quality data from different places, and helps identify areas with poor air quality. This makes indoor air monitoring smarter, more flexible, and more accurate.
We developed a lightweight dual-camera system that detects strawberries and intelligently assigns targets to the appropriate camera, minimizing unnecessary roll-pitch movements. The solution delivers efficient target tracking and accurate control, with only minor deviations due to low-cost hardware limitations.