Breast Mass Classification with Deep Transfer Feature Extractor Model and Random Forest Classifier

Abstract

Breast Cancer is the most common type of cancer & leading cause of deaths in women worldwide. Early diagnosis of breast cancer and proper treatment play a vital role in death rate reduction. The success of Deep Convolutional Neural Networks (CNN) models in image classification tasks with state of art level accuracy has always attracted researchers to use them for disease diagnosis in the field of medical imaging. The proposed method uses CNN based fixed feature extraction technique, a type of deep transfer learning approach to perform binary classification of breast masses using Mammography Images. Mammography images are obtained from three publicly available datasets namely Mammographic Image Analysis Society (MIAS), Digital Database for Screening Mammography (DDSM) and Inbreast. The pretrained CNN models: VGG16, VGG19 & Resnet-50 performs feature extraction from the mammography images. The extracted features from CNN models are then classified into malignant & benign masses using Random Forest machine learning classifier. The models performances have been summarized with performance matrices (Sensitivity, Specificity, F1-score, and Accuracy), onfusion matrices and ROC (Receiver Operating Characteristics) curves. The combination of pretrained models, VGG16/VGG19/Resnet-50 & RF classifier gave model accuracies of: 0.81%, 0.80%,0.83% for the MIAS dataset, 0.994%, 0.986%, 0.996% for DDSM Dataset and 0.83%, 0.81%,0.87% for Inbreast dataset respectively. Automated classification of breast mass from the mammography images can be used by the doctors as a quick and efficient method for breast cancer screening.