Abstract
The reliable detection of diseases using deep learning-based medical image fusion (DLMIF) is a common practice nowadays. The performance of DLMIF depends on the features chosen for the fusion weight calculation. In this study, we examine the efficacy of convolutional neural network (CNN) features for DLMIF utilizing two medical images as input and the fused image produced by several conventional techniques. We use the pre-trained networks in other tasks to extract the feature because DLMIF lacks the ground-truth images necessary to train an end-to-end CNN. The choice of the network and the choice of the convolution layer (CL) are both investigated. We compute consistency maps and the local visibility using the extracted CNN feature map to determine the weight map for DLMIF. The suggested approach performs well with many medical imaging methods. It displays competing quantitative metrics and offers enhanced DLMIF outputs for a better image to be employed in medical diagnosis.