Deep Learning-Enhanced Ultrasound Analysis: Classifying Breast Tumors Using Segmentation and Feature Extraction

Breast cancer remains a significant global health challenge, requiring accurate and effective diagnostic methods for timely treatment. Ultrasound imaging is a valuable diagnostic tool for breast cancer because of its affordability, accessibility, and non-ionizing radiation properties. This study proposes a classification method for breast ultrasound images that integrates segmentation and feature extraction. Initially, ultrasound images are pre-processed to enhance quality and reduce noise, followed by segmentation using the U-Net++ architecture. Feature extraction is then performed using MobileNetV2, and these features are used to train and validate classification models to differentiate between malignant and benign breast masses. The model's performance is assessed using accuracy, precision, recall, Mean IoU, and Dice Score metrics. The U-Net++ model achieved superior segmentation performance with a Dice Score of 0.911 and a Mean IoU of 0.838, outperforming related methods such as U-Net (0.888 Dice, 0.79 IoU) and Efficient U-Net (0.904 Dice, 0.80 IoU). In the classification task, MobileNetV2 when paired with the ANN classifier, produced the highest test accuracy at 96.58%, with a precision of 97% and recall of 96%. Our approach demonstrates superior performance compared to other models, such as RMTL-Net, which achieved 91.02% accuracy, and hybrid CAD models with 94% accuracy. This highlights the benefits of combining advanced segmentation and feature extraction techniques, with MobileNetV2 proving to be the better model, offering superior accuracy and robustness in classification tasks. This approach has the potential to support promise for supporting radiologists, enhance diagnostic accuracy, and ultimately improve outcomes for breast cancer patients. In the future, we will use comprehensive datasets to validate our methodology.