Transfer Learning Assisted Classification of Artefacts Removed and Contrast Improved Digital Mammograms

Abstract

Mammograms are essential radiological images used to diagnose breast cancer well in advance. However, an accurate diagnosis also depends on the quality of mammogram images. Therefore, removal of artefacts and mammogram enhancement are necessary pre-processing steps. Artefact removal helps exclude unsolicited regions in the mammograms and limits the search for suspicious regions without excessive impact from the background. Mammogram enhancements improve apparent visual details and improve some features of an image. In this paper, we propose a method for mammogram pre-processing. These pre-processed mammograms are then fed into Deep Convolutional Neural Network for the classification process. Two approaches are used and compared to classify mammograms; Training model from scratch and Transfer Learning. Transfer Learning is an excellent approach to dealing with the small-sized training set, allowing us to consume the extendibility of deep learning entirely. By employing VGG16 as a pre-trained network on the pre-processed MIAS dataset, we improved training accuracy (96.14\%) compared to the model developed from scratch and other strategies described in the literature.