A Novel Deep Feature Extraction Engineering for Subtypes of Breast Cancer Diagnosis: A Transfer Learning Approach

Abstract

Feature extraction from histological images is a challenging part of computer-aided detection of breast cancer. For this research, we present a novel technique for deep feature extraction for breast cancer diagnosis subtypes based on a transfer learning approach using the BreaKhis dataset. This approach consists of five phases: feature extraction, concatenation, transformation, selection, and classification. In the first phase, nineteen pre-trained convolutional neural networks were used as feature extractors to extract features from the input images. A Support Vector Machine was used at the feature extraction phase to calculate the misclassification rate of each feature generated by the pre-trained networks used. The feature extraction results showed that the two networks achieved the highest accuracy on the dataset and outperformed the other networks. The two networks considered were selected and connected to create the DRNet model, combining the pretrained networks ResNet50 and DenseNet201. The extracted features were decomposed into five sub-hand low-level features using a multilevel discrete wavelet transform in the transformation phase. An iterative neighborhood component analyzer was used to select the minimum number of features needed in the classification phase. A cubic support vector machine was used as a classifier in the final phase. Average classification accuracy of 98.61%, 98.04%, 97.68%, and 97.71% for the 40x, 100x, 200x, and 400x magnification levels, respectively, was achieved.