Transfer Learning Based Approaches for Breast Cancer Classification using Mammogram Images

Breast cancer is a growing epidemic and a leading cause of death among women worldwide. Mammographic imaging has been found to be highly effective in detecting breast cancer at an early stage which leads to reduce mortality rates through prompt and appropriate treatment. In this research work, the proposed models have used two convolutional neural network (CNN) architecture known as VGG19 and ResNet50, which had been pre-trained with data from imageNet and then Mammographic Image Analysis Society (MIAS) database has been used to train and test. To identify potential cancer hotspots, the mammogram images from MIAS database have went through some image preprocessing steps such as image resizing and augmentation by rotation. The extracted features from the pretrained architecture have been flattened into one dimension and then used as inputs to the trainable dense layers. The images have been classified either into benign or malignant type by using the sigmoid activation function. Measures of performance such as accuracy, recall, precision and F1-score have been calculated to evaluate the proposed models’ efficacy. The experimental result depicts that pretrained VGG19 architecture performed 98.46% outperforming ResNet50 achieving a test accuracy of 97.94% .