New Approach to Support the Breast Cancer Diagnosis Process Using Frequent Pattern Growth and Stacking Based on Machine Learning Techniques

Abstract

Breast cancer is one of the most common types of cancer in women, and its early detection significantly improves the survival rate. Although mammography is one of the least invasive and most widely used methods in the diagnostic process, its complexity and subjectivity in medical interpretation present significant challenges. In this article, we propose a new approach that supports the breast cancer diagnosis process by assisting in the classification of mammography images as malignant or benign, or through the BIRADS system. Our proposal consists of two phases. Initially, we implemented the FP-Growth algorithm on patients’ clinical data, analyzing variables such as age and sex to identify frequent patterns. This allows us to explore, group, and visually characterize shared findings and trends among clinical data, which is useful for doctors when creating risk groups or establishing a pre-diagnosis based on the patient’s profile. In this phase, we also prepared the images for training the different models. Subsequently, we combined the strengths of two models through stacking: the Random Forest (RF) model and Convolutional Neural Networks (CNN) with knowledge transfer, to improve image classification and diagnosis. We also explored other methods such as CNN and Support Vector Machine (SVM) to compare the accuracy of the proposed methodology against conventional techniques. The developed models were trained using public datasets: “The Chinese Mammography Database” [2] and “The INbreast database” [3]. The accuracy of the method is evaluated using various classification-related metrics, such as Accuracy, Precision, F1 Score, and Recall. The results show that combining base models using a stacking strategy achieves significantly superior performance compared to individual models, with ideal scores in accuracy, recall, and F1 score using k-fold cross-validation in the meta-model. These excellent results suggest that combining multiple base models more effectively captures the underlying complexities and patterns in the data.