Zhou Yaoyong, Zhu Xiaoliang, Zhou Panyun, Xu Zhenwei, Liu Tianliang, Li Wangjie, Ge Renxian
School of Software, Xinjiang University, Urumqi 830000, China.
Guangdong Polytechnic Normal University, Guangzhou 510000, China.
Curr Med Imaging. 2024;20:e15734056273589. doi: 10.2174/0115734056273589231226052622.
At present, there are some problems in multimodal medical image fusion, such as texture detail loss, leading to edge contour blurring and image energy loss, leading to contrast reduction.
To solve these problems and obtain higher-quality fusion images, this study proposes an image fusion method based on local saliency energy and multi-scale fractal dimension.
First, by using a non-subsampled contourlet transform, the medical image was divided into 4 layers of high-pass subbands and 1 layer of low-pass subband. Second, in order to fuse the high-pass subbands of layers 2 to 4, the fusion rules based on a multi-scale morphological gradient and an activity measure were used as external stimuli in pulse coupled neural network. Third, a fusion rule based on the improved multi-scale fractal dimension and new local saliency energy was proposed, respectively, for the low-pass subband and the 1st closest to the low-pass subband. Layerhigh pass sub-bands were fused. Lastly, the fused image was created by performing the inverse non-subsampled contourlet transform on the fused sub-bands.
On three multimodal medical image datasets, the proposed method was compared with 7 other fusion methods using 5 common objective evaluation metrics.
Experiments showed that this method can protect the contrast and edge of fusion image well and has strong competitiveness in both subjective and objective evaluation.
目前，多模态医学图像融合存在一些问题，如纹理细节丢失导致边缘轮廓模糊，以及图像能量损失导致对比度降低。
为解决这些问题并获得更高质量的融合图像，本研究提出一种基于局部显著性能量和多尺度分形维数的图像融合方法。
首先，利用非下采样轮廓波变换将医学图像分解为4层高通子带和1层低通子带。其次，为融合第2至4层的高通子带，将基于多尺度形态学梯度和活跃度度量的融合规则作为脉冲耦合神经网络的外部刺激。第三，分别针对低通子带和最接近低通子带的第1层高通子带，提出基于改进多尺度分形维数和新局部显著性能量的融合规则，对各层高通子带进行融合。最后，对融合后的子带进行非下采样轮廓波逆变换，得到融合图像。
在三个多模态医学图像数据集上，使用5种常用的客观评价指标，将所提方法与其他7种融合方法进行比较。
实验表明，该方法能很好地保护融合图像的对比度和边缘，在主观和客观评价中均具有较强的竞争力。
