Two-Stage Approach of Hierarchical Deep Feature Representation and Fusion Frameworks for Brain Image Analysis

Abstract

In recent decades, magnetic resonance (MR) brain images have initiated a wide range of image classification and segmentation methods. Feature representation is one of the essential aspects of medical image analysis. This paper proposes and investigates specific features that address the significance of high-level tasks with little annotation for medical images. Deep learning is a futuristic area of research in biomedical image analysis, in which the scope is moving us immediately to the goal of automating tasks in intelligent retrieval systems. This approach can incorporate many levels of feature representation to construct recognition of medical cells or images. We propose a novel approach based on the deep hierarchical features of two different convolutional neural network (CNNs) model choices to achieve competitive performance in the classification task. We explore feature representation through discriminative CNN models. The principal study of our proposed work is feature representations, feature-level fusion, and classification. Meanwhile, effective fusion frameworks were employed for brain MR image classification by using serial fusion and fusion operator strategies. The accuracy of the proposed technique is demonstrated using the Cancer Imaging Archive (TCIA) and Information eXtraction from Images (IXI) datasets. To the best of our knowledge, experiment results show that CNNs feature maps as input to the classifier and are superior to the original CNNs. The performance of the support vector machines (SVM) classifier has been used to evaluate in terms of training performance and classify subjects as either normal or abnormal.