Data Augmentation of Minority Class with Transfer Learning for Classification of Imbalanced Breast Cancer Dataset Using Inception-V3

In this paper, deep learning based experiments are conducted to investigate the effect of data augmentation on the minority class for the imbalanced breast cancer histopathology dataset (BREAKHIS). Two different pre-trained networks are fine-tuned with the minority-augmented dataset. The pre-trained networks were already trained on the well-known ImageNet dataset comprising of millions of high resolution images belonging to multiple object categories. The model so trained is further subjected to transfer learning, to correctly classify cancerous pattern from non-cancerous conditions, in a supervised manner. Our experiments were carried out in two phases. Phase-I investigates the effect of data augmentation applied on minority class for the Inception-v3 and ResNet-50 pre-trained networks. Results of phase-I are further enhanced in phase-II by the transfer learning approach in which features extracted from all layers of Inception-v3 are learnt by the SVM and weighted SVM classifiers. From experimental results, it was found that the pre-trained Inception-v3 model with data augmentation on minority class outperforms other network types. Results also indicate that Inception-v3 with data augmentation of minority class and transfer learning with weighted SVM gives the highest classification accuracies.