Beyond Boundaries: A Novel Ensemble Approach for Breast Cancer Detection in Ultrasound Imaging Using Deep Learning

Breast cancer remains a major global health concern, necessitating advanced methods for accurate detection and classification. This research presents an innovative approach using an Enhanced Deep Neural Network (EDNN) model, which outperforms the Vision Transformer and traditional transfer learning methods. Combining the MobileNet and Xception models makes our EDNN more accurate in identifying and classifying breast cancer. As part of our preprocessing, we scale images, standardize data, and apply data augmentation to prepare the dataset for analysis, enhancing the model's format and generalizability. The EDNN, widely used for breast cancer imaging, is trained and validated using ultrasound images. The EDNN surpasses existing transfer learning models, such as the Vision Transformer, with accuracy rates of 97.89% and 96.86% on Datasets 1 and 2, respectively. The AUC value of 0.98 on Dataset 1 underscores the model's robustness. The Grad-CAM Explainable Artificial Intelligence (XAI) method enhances the interpretability of our model, making it more transparent and understandable. Additionally, we employ the U-Net segmentation method to separate images, enabling precise identification and isolation of specific areas of interest to examine breast cancer features thoroughly. Our research, leveraging state-of-the-art deep learning algorithms and image segmentation techniques, significantly advances the effectiveness of breast cancer detection and treatment.