基于误差反馈的多尺度特征网络的CT图像超分辨率重建

CT Image Super-resolution Reconstruction Based on Error Feedback Multi-scale Feature Network

目前,基于多尺度特征卷积神经网络的医学图像超分辨率重建算法取得了良好的重建效果,但是这些方法还面临着多尺度特征类型单一,特征提取与融合中易丢失高频信息以及模型规模过大的问题.针对以上问题,提出一种基于误差反馈的多尺度特征网络模型的医学图像超分辨率重建方法.该方法使用不同深度的并行卷积层实现感受野的区别,设计了一种基于误差反馈的多尺度特征提取模块,从而实现多尺度特征的提取.该方法还使用全局反馈结构逐步提取高频信息,最后使用误差反馈反向传递高频信息到每次循环提取的特征中,利用融合后的特征重建出高分辨率CT图像.实验结果表明:(1)该方法在×2,×4,×8这3种放大倍率上都取得了良好的效果,与主流的超分辨率算法相比,在客观指标和主观评价上取得了更好的结果;(2)该方法重建的CT图像,更清晰,细节纹理更丰富.

At present,medical image super-resolution reconstruction algorithms based on multi-scale feature convolutional neural networks have achieved good reconstruction results,but these methods still face the problems of the single type of multi-scale features,easy loss of high-frequency information in feature extraction and the fusion,and the excessive model size.To address these problems,this paper proposes a super-resolution reconstruction method for medical images based on error feedback multi-scale feature network model.The method uses parallel convolutional layers of different depths to achieve perceptual field differentiation,and an error feedback-based multi-scale feature extraction module is designed to enable the extraction of multi-scale features.The method also uses a global feedback structure to gradually extract high-frequency information,and finally uses error feedback to reverse the high-frequency information to the features extracted in each cycle,and uses the fused features to reconstruct a high-resolution image.The experimental results show that(1)the method achieves good results for all three magnifications×2,×4,×8,achieving better results in terms of objective metrics and subjective evaluation compared with the mainstream super-resolution algorithms;(2)the CT images reconstructed by the method are clearer and richer in detail texture.