Abstract

Brain tumor (BT) is a serious cancerous disease caused by an uncontrollable and abnormal distribution of cells. Recent advances in deep learning (DL) have helped the healthcare industry in medical imaging for the diagnosis of many diseases. One of the major problems encountered in the automatic classification of BT when using machine learning (ML) techniques is the availability and quality of the learning from data; these are often inaccessible, very confidential, and of poor quality. On the other hand, there are more than 120 types of BT [1] that we must recognize. In this paper, we present an approach for the automatic classification of medical images (MI) of BT using image fusion (IF) with an auto-coding technique for data augmentation (DA) and DL. The objective is to design and develop a diagnostic support system to assist the practitioner in analyzing never-seen BT images. To address this problem, we propose two contributions to perform data augmentation at two different levels: before and during the learning process. Starting from a small dataset, we conduct the first phase of classical DA, followed by the second one based on the image fusion technique. Our approach allowed us to increase the accuracy to a very acceptable level compared to other methods in the literature for ten tumor classes.