Regresar

Breast Cancer Screening Using Deep Learning

Abstract:

Breast cancer is a disease in which the cells lining the ducts or lobules of the breast glandular tissue begin to grow out of control. Early detection increases the probability of being overcome by the patient. Breast cancer screening seems to be one of the best ways to provide early care for these patients, especially those who are asymptomatic. On the other hand, it has been indicated that between 10-30% of breast cancer studies are misclassified, mainly due to errors in interpretation (52% ), errors in image search (43%), and poor technical quality (5% ). Nowadays deep learning techniques have given great results in different medical imaging studies. This paper aims to present an analysis and evaluation of machine learning techniques for (1) screening for the presence of breast cancer and (2) classification of the level of malignancy if present. The deep convolutional neural networks analyzed are VGG-16, VGG-19, MobiINet-V2, and DenseNet-121. These architectures were used as binary classifiers, the model suggests the presence of a malignant tumor or not. K-Nearest-Neighbor (KNN), Random Forest, Gradient Boosting, and AutoML, were evaluated as multi-class classifiers for disease grade classification. The comparative analysis was performed by standardizing the test environment with the same number of epochs and analyzing the values obtained in terms of accuracy, training time, and F1. It was concluded that the DenseNet-121 was the best classifier of the presence of malignancy or not, however, for considering healthy and diagnosed cases MobilNet-V2 showed better results but with a small difference. Random Forest was the best classifier for a grade of malignancy.