基于多维动态卷积的渐进式双分支医学图像融合模型OA

To make the fused image possess both abundant texture details and clear structural features,a multi-dimensional dynamic convolution-progressive dual-branch medical image fusion(MDC-PDF)model was proposed.First,a gradient path and a contrast path based on the difference connection of source images were constructed to enhance the detail information and contrast of the fused images.Subsequently,the MDC module was used instead of traditional convolution to extract more detailed information.Then,the image-wise transfer blocks(ITB)module were introduced into different layers of the fused features,and the fused features were fed back into the two branches to promote information exchange between features.Finally,the contrastive language-image pretraining(CLIP)loss was introduced to guide the training of the network.The results demonstrated that the MDC-PDF model was superior to other comparative models in terms of entropy,spatial frequency,mutual information and visual information fidelity,and that the visual information fidelity was improved by up to 5.03%compared with that of the suboptimal model.High-quality fused images with abundant textures and clear structural characteristics were generated by MDC-PDF model.