基于卷积神经网络模型的医学图像融合

Medical Image Fusion Based on Convolutional Neural Network Model

提出了一种新的基于卷积神经网络(CNN)和加权最小二乘法(WLS)的医学图像融合算法。算法主要步骤如下:利用滚动导向滤波(RGF)和高斯滤波(GF)构成的混合多尺度分解工具将源图像分解为基础层和一系列细节层,从而能够更好地保留尺度信息和边缘信息。基于卷积神经网络给出基础层融合规则,该规则能够更好地提取图像特征,使融合图像能够很好继承源图像结构信息、能量信息和强度信息。利用绝对值取大规则和加权最小二乘法优化策略,对细节层进行融合,使融合图像中包含更多的视觉细节信息和具有更高对比度。实验结果表明所提算法在视觉评价和客观评价方面与其他算法相比具有较好的优势,且在急性中风、致命性中风和脑膜瘤这三类疾病图像融合效果更为突出。

A new medical image fusion algorithm based on convolutional neural network(CNN)and weighted least squares(WLS)is proposed in this paper.The main steps of the algorithm are as follows:the source image is decomposed into a base layer and a series of detail layers using a hybrid multiscale decomposition tool consisting of rolling guided filter(RGF)and Gaussian filter(GF),which can better preserve the scale information and edge information.Based on convolu-tional neural network,the fusion rule of base layer is given,which can extract image features better,and the fused image can inherit structure information,energy information and intensity information of the source image well.Combining the“max-absolute”fusion rule with the weighted least squares optimization strategy,the fusion rule of details layer is given to make the fused image contain more visual details and have higher contrast.The proposed algorithm has better advantages than other methods in visual and objective evaluation,and is more effective in image fusion for acute stroke,fatal stroke and meningioma.