Advanced Multi-Scale Vision Transformer with Neural Architecture Search for Thyroid Cancer Diagnosis Using Ultrasound Images

This project addresses road extraction from very high-resolution satellite images, tackling occlusions, clutter, and thin road shapes. Two segmentation models—DeepAttnLab (with attention mechanisms) and PSPNet (with pyramid pooling)—share a ResNet50 backbone and stride modification to retain spatial detail. A combined BCE, Dice, and Focal loss handles class imbalance. Both models outperform the SDFFNet baseline. The best model is deployed in a Flask web application featuring user login, a relevance classifier to reject non-satellite inputs, and real-time road mask prediction, achieving accurate and continuous road extraction.