Bi-Modal Transfer Learning for Classifying Breast Cancers via Combined B-Mode and Ultrasound Strain Imaging

Detecting breast cancer accurately is still difficult in medical imaging. Doctors need to tell the difference between dangerous (malignant) and harmless (benign) tumors. This study creates a new system that uses two types of ultrasound images together: B-mode images (showing tumor shape and structure) and strain elastography images (showing how stiff or soft the tissue is). Both images must be taken from the same patient and show the same tumor area. The system uses two smart computer programs called AlexNet and ResNet-18, which were first trained on millions of regular photos and then retrained specifically for breast cancer images. The B-mode image (one gray channel) and strain elastography image (three color channels) from the same patient are combined into a single four-channel image. Both computer programs learn from these combined images together. The early layers of both programs are frozen to keep basic image patterns they already learned, while the later layers are trained to recognize breast cancer features. After training, features from both programs are joined together—AlexNet provides detailed texture information while ResNet-18 provides deeper pattern information. A final decision-making layer uses these combined features to classify tumors as malignant or benign. The system shows excellent performance in accuracy, sensitivity, specificity, precision, and F1-score when tested on new images. This approach successfully combines information from both imaging types, giving doctors a reliable computer tool to help detect and classify breast cancer more accurately.

Keywords: Breast cancer classification, bi-modal ultrasound imaging, transfer learning, deep convolutional neural networks, ensemble model