Deep-Learning Model for Breast Tumor Identification Using a Hybrid Approach With Pre-Trained Models

Globally, breast cancer is among the most common types of cancer affecting women, and early detection can save lives. In case of breast cancer, early detection improves the chances of a favorable prognosis since prompt treatment can be commenced. Despite the lack of access to specialist physicians, machine learning facilitates early diagnosis of breast cancer. With the rapid advancement of machine learning, especially deep learning, the medical imaging community is increasingly interested in using these methods to improve cancer detection accuracy. Nevertheless, significant deficiencies persist in the existing literature: primary challenges encompass the restricted availability of extensive, balanced, and meticulously annotated datasets; complications regarding model generalizability and external validation; and enduring class imbalance that may result in false negatives and affect clinical reliability. Moreover, there is an absence of established standards and challenges in comparing studies owing to different datasets and evaluation processes. Many deep learning models exhibit robust performance on internal test sets; however, they frequently struggle to sustain accuracy when evaluated on external, heterogeneous data, which raises questions regarding their practical applicability. A limited amount of data is available on diseases. The objective of this study is to present a deep learning model for identifying and classifying breast cancer. Kaggle Open-Access Database breast histopathology images were used to assess the system's performance. Several quantitative criteria are used to evaluate the efficacy of the suggested strategy, including accuracy, precision, recall, and the F1 score. All previous models fail to match the performance of the ensemble model, composed of VGG16, EfficientNetB7, and Xception. As result of the study, the accuracy, precision, recall, and F1-scores for invasive ductal carcinoma (IDC) are all 95.1%, 96.5%, and 95.77% respectively. The results emphasize the strength and dependability of our innovative methodology in automating breast cancer classification, surpassing previous research endeavors, optimizing diagnostic procedures, and ultimately helping to preserve the lives of patients.