Classification of Breast Cancer Histopathological Images Using KAZE Features

Breast cancer (BC) is a public health problem of first importance, being the second most common cancer worldwide. BC represents 30.4% of all new cancer cases in the European female population. The diagnosis and differential diagnosis of BC is based on the clinical presentations, physical examinations combined with imaging studies, and confirmed by histopathologic findings. Pathologists' examination is a time-consuming analysis, susceptible to an interpretation bias mainly caused by the experience of the pathologist and the decrease of attention due to fatigue. Currently, computer-aided detection and diagnosis techniques applied to digital images are assisting the specialists. In this work, the performance of a pattern recognition system based on KAZE features in combination with Bag-of-Features (BOF) to discriminate between benign and malignant tumours is evaluated on the BreakHis database (7909 images). During the training stage, KAZE keypoints are extracted for every image in the training set. Keypoints are mapped into a histogram vector using K-means clustering. This histogram represents the input to build a binary SVM classifier. In the testing stage, the KAZE keypoints are extracted for every image in the test set, and fed into the cluster model to map them into a histogram vector. This vector is finally fed into the binary SVM training classifier to classify the image. The experimental evaluation shows the feasibility and effectiveness, in terms of classification accuracy, of the proposed scheme for the binary classification of breast cancer histopathological images with a low magnification factor.