Medical Image Fusion through Siamese Network-Based Spatial Domain Fusion with Gaussian Pyramid Decomposition

Abstract

Spatial domain-based medical image fusion methods were among the earliest research focal points. However, spatial domain techniques tend to introduce spectral and spatial distortions into fused images. This research introduces the Siamese network, one of the three models for comparing patch similarity in convolutional neural networks (CNN). Its two identical weight branches ensure uniform feature extraction and activity level measurements, offering advantages over pseudo-Siamese and 2-channel models. The ease of training makes the Siamese model a preferred choice in fusion applications. The proposed method utilizes weight maps, Gaussian pyramid decomposition, and pyramid transform for multiscale decomposition, aligning the fusion process with human visual perception. Additionally, a localized similarity-based fusion strategy is employed to adaptively adjust decomposed coefficients. This algorithm combines pyramid-based and similarity-based fusion techniques with CNN models, resulting in an advanced fusion approach.