Analysis of breast cancer classification robustness with radiomics feature extraction and deep learning techniques

The diagnostic of breast cancer and breast imaging procedures are typically carried out using a variety of imaging modalities, including mammography, MRI, and Ultrasound. However, Ultrasound and mammography have limitations. MRI is better than other procedures. Recent computational approaches, such as the Radiomics applied to image analysis, have shown remarkable progress for removing diagnostic difficulties. This thesis analyzed the robustness of breast tumor classification with features extraction (radiomics) and featureless method (deep learning). It contains two stages: the first stage introduced and explored radiomics based steps. A total of 111 tumor lesions were used to derive 74 radiomic features consisting of shape, first-order, and three separate second-order metrics. Four separate associations of features were used to classify tumor lesions with four different kernels from support vector machine algorithm. Second-order defined data split showed better cross-validation performance with highest accuracy of 96.17%, where all feature combinations data split showed 96.08% accuracy. In the confusion matrix analysis, the SVM-RBF kernel developed optimal diagnostic efficiency with a maximum test accuracy of 97.06% on two separate combination data group analysis. The second stage developed with deep learning techniques (InceptionV3 and CNN-SVM). A total of 2998 images were used to create the models. In this portion, the CNN-SVM model achieved the highest accuracy, 95.28%, with an AUC of 0.974, where the pre-trained InceptionV3 achieved an AUC of only 0.932. Finally, the obtained result in both stages was discussed together and other related studies. Keywords: Breast Tumor Classification, Radiomic Features, Deep Learning.