基于Restormer与双重注意力机制的非对称医学图像融合模型

打开文本图片集

关键词：医学图像融合；双重注意力机制；非对称融合；两阶段训练 中图分类号：TP391 文献标识码：A doi：10.37188/CJLCD.2025-0087 CSTR：32172.14.CJLCD.2025-0087

Asymmetric medical image fusion model based on Restormer and dual attention mechanism

KONG Weiwei，LI Zejiang，HE Leilei， DU Yusheng （School of Computer Science and Technology， Xi'an University ofPosts & Telecommunications， Xi'an 710121，China）

Abstract：There are differences in spatial information distribution among diferent medical imaging models， which is not conducive to the efective alignment of the depth feature space，resulting in the loss of shallow information in a specific area of the fusion image or excessive dependence on the information ofa certain mode.To solve these problems，an asymmetric medical image fusion model based on Restormer and dual attention mechanism was proposed.Firstly，Restormer module is used to dig deep features of different modal images，and dual attntion mechanism is introduced to extract global and local features of different modalimages.Secondly，anasymmetric feature fusion strategy isdesigned，in whichan independent feature encoderis designed for each mode and the extracted features are fused.Finally，the fused features are generated by the decoder. This model adopts two stages of training，the first stage mainly extracts global and local features from diffrent modal images，and attempts to reconstruct the original image to calculate the loss；the second stage continues to extract deep features and generate fusion images. Compared with the seven mainstream image fusion models，the seven evaluation indicators，standard deviation，spatial frequency， visual information fidelity，spectral relevance，mutual information，average gradient，and ΔQ index used to evaluate hybrid fusion have an average increase of 12.63% ，28. 30% ， 31.37% ， 27.40% ， 19.01% ， 37.36% ， 32.44% ，respectively. The fusion strategy of this model can not only efficiently integrate the coding features from diffrent modes，but also complete the integration of complementary information and the interaction of global information without manually designing fusion rules，and can better integrate images from different modes.

Key words： medical image fusion； dual attention mechanism； asymmetric fusion；two-stage training

1引言

随着计算机视觉技术的持续革新与传感器设备的不断进步，图像融合技术作为突破单模态感知局限的一个重要课题也在不断地发展，其目的是通过结合源图像中的重要信息来生成信息丰富的融合图像[1]。（剩余17612字）