The diagnostic accuracy of deep learning models in medical imaging heavily relies on the consistency and quality of input data. However, ultrasound images can vary significantly due to differences in scanner manufacturers, imaging protocols, and operator techniques, posing a challenge for model generalizability across domains. This project investigates the migration performance of multiple state-of-the-art deep learning architectures—DenseNet-201, ResNet-34, and MobileNetV3 enhanced with an Iterative Local Normalization Layer (IterLNL)—on breast ultrasound image classification. A robust ensemble model is also constructed to leverage complementary strengths of individual networks. Using the BUSI dataset, these models are evaluated on a stratified test set and assessed via classification metrics, confusion matrices, and accuracy comparisons. Results indicate that while individual models perform competitively, the ensemble approach consistently yields superior generalization performance. This study highlights the significance of model ensemble and domain-aware architectural adaptations for robust deployment across heterogeneous ultrasound systems. Our findings emphasize the need for transfer-aware deep learning pipelines to ensure reliability in real-world, scanner-diverse clinical environments.

Keywords: Breast Ultrasound Imaging, Deep Learning,Model Migration,Transfer Learning,DenseNet-201,ResNet-34,MobileNetV3,Iterative Local Normalization Layer (IterLNL),Ensemble Learning,Medical Image Classification,Generalization Across Devices,Multi-scanner Robustness,Diagnostic AI,Confusion Matrix Analysis