A Hybrid Approach for Breast Mass Categorization

Breast cancer is one of the most frequent fatal diseases among women around the world. Early diagnosis is paramount for easing such statistics, increasing the probability of successful treatment and cure. This paper proposes a hybrid approach composed of a convolutional neural network with a supervised classifier on the top capable of predicting eight specific cases of the breast tumor, being four of them malignant and four benign. The model employs the BreastNet convolution neural network to the task of mammogram images feature extraction, and it compares three distinct supervised-learning algorithms for classification purposes: (i) Optimum-Path Forest, (ii) Support Vector Machines (SVM) with Radial Basis Function, and (iii) SVM with a linear kernel. Moreover, since BreastNet is also capable of performing classification tasks, its results are further compared against the other three techniques. Experimental results demonstrate the robustness of the model, achieving \(86\%\) of accuracy over the public LAPIMO dataset.