Detection Of Breast Cancer From Mammography Images

Abstract

Early detection is very important in minimizing mortality rates; however, the manual diagnosis using mammogram images is complicated and requires expert intervention. Several AI-based approaches have been reported in the literature, but they still suffer from several challenges, such as poor feature extraction, inadequate training models, and the inability to differentiate between malignant and non-cancerous regions.  We introduce a new automatic computational framework for breast cancer classification in this paper. This model employs a novel technique known as haze-reduced local-global contrast to enhance image contrast. The enhanced images are then used for dataset augmentation to increase the range of datasets and enhance the training effectiveness of the selected deep learning model. Then, a pre-trained model named EfficientNet-b0 was used and augmented with extra layers.   The enhanced model was separately trained on both the original and improved images with deep transfer learning methods and fixed hyperparameter initialization.  In the second phase, a new serial-based method was employed to combine deep features that had been gathered from the average pooling layer.