Multi-Modal Medical Image Fusion Using 3-Stage Multiscale Decomposition and PCNN with Adaptive Arguments

Abstract

In the current era of technological development, medical imaging plays an important role in many applications of medical diagnosis and therapy. In this regard, medical image fusion could be a powerful tool to combine multi-modal images by using image processing techniques. But, conventional approaches failed to provide the effective image quality assessments and robustness of fused image. To overcome these drawbacks, in this work three-stage multiscale decomposition (TSMSD) using pulse-coupled neural networks with adaptive arguments (PCNN-AA) approach is proposed for multi-modal medical image fusion. Initially, nonsubsampled shearlet transform (NSST) is applied onto the source images to decompose them into low frequency and high frequency bands. Then, low frequency bands of both the source images are fused using nonlinear anisotropic filtering with discrete Karhunen–Loeve transform (NLAF-DKLT) methodology. Next, high frequency bands obtained from NSST are fused using PCNN-AA approach. Now, fused low frequency and high frequency bands are reconstructed using NSST reconstruction. Finally, band fusion rule algorithm with pyramid reconstruction is applied to get final fused medical image. Extensive simulation outcome discloses the superiority of proposed TSMSD using PCNN-AA approach as compared to state-of-the-art medical image fusion methods in terms of fusion quality metrics such as entropy (E), mutual information (MI), mean (M), standard deviation (STD), correlation coefficient (CC) and computational complexity.

Adaptive convolutional sparsity with sub-band correlation in the NSCT domain for MRI image fusion

“…Extensive research has been carried out in recent years on the topic of multimodal image fusion. According to the image transformation strategy adopted in Li et al (2016), fusion categories are mainly concentrated on four aspects: MST-based schemes (Tannaz et al 2020, Goyal et al 2021, Singh et al 2022, sparse representation (SR)based schemes (Yin 2018, Shibu and Priyadharsini 2021, Zhang 2021, deep learning (DL)-based schemes (Vasanthi et al 2021, Liu et al 2022, and the combination of different transformations (Maqsood and Javed 2020, Vanitha et al 2020, Reddy et al 2021. It is worth noting that the applications of SR in voxel selection, compressed sensing, and bioelectrical signal detection prove that SR has become an effective tool for image processing.…”

A multimodal molecular image fusion method based on relative total variation and co‐saliency detection

“…The most widely used are the non-subsampled contourlet transform 31 and the non-subsampled shearlet transform. 32 They can flexibly decompose the input image into different directions as well as different scales. And the computational complexity increases with the degree of the decomposition.…”

“…In addition, other types of tools further break through existing limitations. The most widely used are the non‐subsampled contourlet transform 31 and the non‐subsampled shearlet transform 32 . They can flexibly decompose the input image into different directions as well as different scales.…”