Abstract:

For researchers to improve treatment efficacy and reduce mortality rates, early identification of breast cancer is crucial assisted by digitally enabled tools. In recent times, deep CNN-based deep transfer learning (TL) methods have emerged as the most facilitating technology. In this paper, we propose a computer aided methodology named deep hybrid AS-RF-ShuffNetV2 which addresses three major challenges of the base technology: data imbalance, extraction of feature set, and classification. Three steps make up the entire work: (1) Adaptive synthetic minority oversampling (ADASYN) is used to oversample malignant images (minority) in order to improve the feature space; (2) Random Forest (RF) splits are used to extract selective features from the balanced dataset; and (3) ShuffleNetV2, an efficient deep TL model, uses a channel split block and a decisive channel attention block to classify the final features into two target classes. The accuracy and AUC score achieved by our model are 92.68% and 0.86, using INBreast dataset. The strength and consistency of the suggested method in correctly identifying breast cancer classes are demonstrated by thorough testing versus existing contemporary approaches.