Fusion of MRI Images and PET Images Based on ResNet50

When performing image fusion, the resulting fused image directly depends on the model used for fusion. Different fusion models result in different merits of the fused image. Knowing the feature extraction and processing methods of the source image is the core technology of fusion. Most of the deep learning-based methods are still inadequate in the way they extract and process the features of the source image. This often leads to degradation of the fusion performance in some cases, thus failing to achieve the expected requirements. To address this issue, we use a novel fusion framework based on Resnet50. At first, we extract depth features from the input image using a residual network (ResNet), then normalize the extracted depth features using phase component analysis (ZCA) and I1-norm and obtain an initial weight map, then use the soft-max operation associated with the initial weight map to obtain our final weight map, and finally reconstruct the fused image using a weighted averaging strategy, compared with the existing fusion methods, the fusion framework we use visually better fuses the depth features of the source image.