Grad-CAM Guided Preprocessing and Convolutional Neural Network for Efficient Mammogram Images Classification

Abstract

Breast cancer is now widely recognized as the second most deadly disease in women. Computer Aided Diagnosis system (CAD), and specifically deep learning (DL), have continued to provide a significant computational solution for early detection and diagnosis of this disease. Research efforts are advancing new approaches to improve the performance of deep learning-based models. We are developing a new system (CAD) to classify mammograms as normal or abnormal and then classify the abnormal ones as benign or malignant. The proposed system consists of three main steps. The pre-processing step consists mainly in cropping the images with the application of Grad-CAM (Class Activation Map) method which generates a heatmap that facilitates the detection of the area of interest to be cropped and minimizes the pre-processing steps usually done in previous works. A contrast enhancement is performed with the Contrast Limited Adaptive Histogram Equalization (CLAHE) method to differentiate the distinct elements of the image. We then proceed to data augmentation with two different ways combining the usual geometric transformations such as rotations, shifts and translations, before proceeding to the training of the different proposed models, namely: VGG16, Vgg19, Resnet50, Densenet121 and InceptionV3, thanks to which we achieved the highest accuracy of 99.13% for the benign and Malignant classification using the pre-trained network Resnet50, and an accuracy of 98.54% with the pre-trained network Vgg19.