MIGRATION OF DEEP LEARNING MODELS ACROSS ULTRASOUND SCANNERS

Abstract

This project presents a deep learning framework to address domain shift in ultrasound imaging by enabling migration of models across different scanners. The study utilizes the BUSI (Breast Ultrasound Images) dataset, which includes images categorized into benign, malignant, and normal classes. The proposed approach combines EfficientNet-B0 as a feature extraction backbone with a Domain-Adversarial Neural Network (DANN) to learn domain-invariant features, supported by a Gradient Reversal Layer for effective adaptation between source and target domains. The model is trained using supervised learning on the source domain while minimizing domain discrepancy with adversarial training. Experimental results demonstrate strong performance, achieving approximately 90% accuracy along with consistent F1- scores across all classes, indicating robust generalization. The loss function integrates classification and domain loss, and during training, the overall loss converges to a very low range (approximately 0.0 to 0.1), reflecting stable and efficient learning. Overall, the combination of EfficientNet-B0 and DANN significantly improves crossscanner reliability, making the model suitable for real-world clinical deployment.