Automatic Histopathological Image Classification using Multi-Deep Features with Handcrafted Features and Shallow Machine Learning Models

Invasive ductal carcinoma (IDC) is one of the most common variant of breast cancer affecting women more frequently than any other cancer. One crucial clinical duty is to correctly detect and classify breast cancer subtypes. Automated techniques can help to reduce errors and save time. However, determining the malignancy from tissue biopsies is difficult and dependent on the subjectivity of the observer. To diagnose breast cancer, a novel classification model using fusion of deep features VGG16 and VGG19 with handcrafted feature GLCM alongside the Medium Gaussian SVM classifier is developed in this paper. Additionally, thirteen hybrid classifiers are used to study the classification of breast cancer. These classifiers use four deep features, three handcrafted features, and a further fusion of deep features to produce six more combinations. Resnet50, VGG16, VGG19 and GoogleNet (Inception v3) are the pre-trained networks used to extract the deep features. Similarly, the handcrafted feature extraction techniques used are HOG, LBP and GLCM. Following evaluation, the classification model is framed by considering the best-performing deep features (VGG16 and VGG19 fusion model), handcrafted feature (GLCM) and classifier (Medium Gaussian SVM). Area Under Curve (AUC) of 0.91, Accuracy of 87.8% and Positive Likelihood Ratio (LR+) of 12.2 were attained by the Medium Gaussian SVM based classification model that combined the features of GLCM with the fusion model of VGG16 and VGG19. According to STARD guidelines, the recommended diagnostic technique is Excellent (AUC is between 0.9 and 1.0 and LR+ is above 10.0). The achieved results also outperformed the recently published articles on the same domain tackling similar problems.