A hybrid approach for classification Breast Cancer histopathology Images

Abstract

Breast cancer is a significant factor in female mortality. Automated identification and classification of breast histopathology image tissue characteristics using computer-aided diagnostic tools is an important step in disease identification and therapy. In this work we propose an automated classification system, which is based on mixing pre-trained deep Convolutional Neural Networks (CNN) as feature extractor, and multilevel hand-crafted features. The pre-training model is used: ResNet18, Inception ResNet v2, ShuffleNet, and Xception. for hand-crafted features are extracted using Haralick textures, Rotation and Scale-invariant Hybrid image Descriptor (RSHD), Local Diagonal Extrema Pattern (LDEP), Speeded up robust features (SURF), Colored Histogram, and Dense Invariant Feature Transform (DSIFT) set All extracted features reduced by feature selection method (PCA) and use them as a feature vector for the training three classifier Support Vector Machine (SVM), Random Forest (RF), and K-Nearest Neighbor (KNN. We evaluate the efficiency of the proposed methodology on publicly microscopy ICIAR-2018 dataset that contains histopathology images to four classes: invasive carcinoma, in-situ carcinoma, benign tumor, and normal tissue. Experimental results show the accuracy of the proposed method between 96.97%.