Medical image fusion via decoupled representation and component-wise regularization learning

Rui Zhang, Haoze Sun, Lizhen Deng, Hu Zhu, Wei Qian

Abstract:

Medical image fusion plays an important role in the precise diagnosis, treatment planning, and follow-up
studies of various diseases. While tremendous improvements in medical image fusion based on convolution
sparse coding have been achieved, existing methods are still limited by the intractable redundancy information
interaction between source medical images. In this paper, we propose an easy yet effective representation
and regularization learning method based on decomposed components scheme with high competitive performance. We construct more compact information interactions by decoupled representation learning, which
simultaneously mitigates the problem of redundancy in fusion component entanglement. And then two different
regularization operators are adaptively exploited to depict two different components separately, which describe
the structural-inspired difference based on the decoupled principle. Furthermore, we combine the alternating
direction method of multipliers (ADMM) algorithm and the conjugate gradient (CG) method to optimize our
proposed model. Our experiments demonstrate that our proposed method has significant improvements in
efficiency and fusion performance against the state-of-the-art methods.