Feasibility of Convolutional Neural Networks (CNN) for the Fusion of Temporal Medical Images

Abstract

In medical applications, a decision is made based on the collective information obtained about an abnormality after analyzing a series of images instead of a single image. Analyzing a set of images and combining the information from each image is time consuming and tedious. Instead, if the features of interest in the set of images could be provided on a single image, decision making becomes easy. Process of combining images is termed as image fusion. This paper reports the image fusion techniques performed in spatial domain (Principal Component Analysis (PCA) based fusion), transform domain (Discrete Cosine Transform (DCT)) and Convolutional Neural Network (CNN) used for the fusion of MRI images. Convolutional layers and max pooling layers in VGG16, VGG19 and ALEXNET are used for extracting the features and fusion rules are used for obtaining the output images. Performance is measured in terms of Signal to Noise Ratio (SNR), Root Mean Square Error (RMSE), entropy and standard deviation. It is observed that VGG19 outperforms other image fusion techniques and provides consistently good performance which is evident from the performance evaluation.