Convolutional neural network based early detection of breast cancer using mammography images

According to World Health Organisation (WHO) data, breast cancer is the most common type of cancer in women and the mortality rate from this type is very high. Accurate and timely detection of breast cancer is a critical factor that directly affects the success of the treatment process. Traditional mammography and imaging methods can vary depending on the experience of radiologists, which can lead to false positive or negative results. In this context, the integration of artificial intelligence and deep learning technologies provides a significant improvement in cancer diagnosis. Deep learning algorithms offer high accuracy rates in recognising abnormalities in mammography images thanks to their ability to process and learn from large data sets. The aim of this study is to develop a new model that improves the accuracy and reliability of deep learning based classification systems in breast cancer diagnosis. The Breast Cancer Ensemble Convolutional Neural Network (BCECNN) model was developed using the AISSLab dataset. The model uses triple (TECNN) and quintuple (FECNN) ensemble structures in which the outputs of five different deep learning architectures (AlexNet, VGG16, ResNet-18, EfficientNetB0, XceptionNet) are combined by majority voting. This methodology aims to obtain more accurate and reliable predictions by combining the strengths of each network. The model is optimised with transfer learning techniques to avoid the overfitting problem and to increase its generalisation capacity. Transfer learning enables the model to obtain better results with limited data, thus contributing to more efficient training. The performance of the BCECNN model was extensively evaluated by creating five different classification subsets from the AISSLab dataset (AISSLab-v1, AISSLab-v2, AISSLab-v3, AISSLab-v4, AISSLab-v5). These tests allowed to analyse the accuracy, reliability and generalisability of the model under different clinical scenarios. The results show that the BCECNN model achieves an accuracy of 98.75% and is one of the most successful models in the current literature. This shows that the model can be a highly effective tool in breast cancer diagnosis and can be used safely in clinical applications. One of the important contributions of the model is the integration of Explainable Artificial Intelligence (XAI) techniques. By using explainability methods such as Gradient Weighted Class Activation Mapping (Grad-CAM) and Locally Interpretable Model Independent Explanations (LIME), the decision processes of the model are made transparent. In this way, clinical experts were able to visualise which image regions were considered critical by the model and better analyse the reliability of the model's decisions. The support of explainable artificial intelligence increases the confidence of clinical experts in the results of the model and facilitates the use of the model in clinical decision support systems. In conclusion, the BCECNN model offers significant advantages in terms of high accuracy, explainable artificial intelligence support and clinical applicability. The model has great potential to contribute to the future use of artificial intelligence-based clinical decision support systems. This study demonstrates how deep learning techniques can be effectively used in critical health problems such as breast cancer diagnosis and how artificial intelligence applications can be made more efficient in the health sector. In this context, the developed BCECNN model constitutes an important step to support the integration of artificial intelligence-based systems into clinical applications in healthcare.