计算机工程与应用 ›› 2022, Vol. 58 ›› Issue (23): 230-237.DOI: 10.3778/j.issn.1002-8331.2105-0202

• 图形图像处理 • 上一篇    下一篇

多重放大的医学图像超分辨率重建

章伟帆,曾庆鹏   

  1. 南昌大学 信息工程学院,南昌 330031
  • 出版日期:2022-12-01 发布日期:2022-12-01

Multi-Scale Medical Image Super-Resolution Reconstruction

ZHANG Weifan, ZENG Qingpeng   

  1. College of Information Engineering, Nanchang University, Nanchang 330031, China
  • Online:2022-12-01 Published:2022-12-01

摘要: 针对当前大多数基于深度学习的医学图像超分辨率重建方法存在放大因子单一的问题,提出一种多重放大的医学图像超分辨率重建网络模型。以密集残差网络为基础,特征提取级联多个改进连接的密集残差块,降低连接复杂度至对数级,避免浅层的医学图像特征被反复处理。特征图重建采用元信息直接嵌入模式,利用一个小型网络学习不同放大因子任务间的通用知识,实现不同重建任务的整合。将不同放大因子任务对齐至同一维度,实现对小数重建任务的支持。实验结果表明,所提方法与深度卷积(VDSR)等典型方法相比,在峰值信噪比(PSNR)与结构相似度(SSIM)上仍有0.17~1.57 dB与0.002 2~0.042 5的提升。

关键词: 超分辨率, 深度学习, 密集残差, 元信息嵌入, 医学图像

Abstract: Aiming at the problem that most prior methods based on deep learning focus on the alone scale factor medical image super-resolution task, a multi-scale super-resolution network for medical image is proposed. The model is optimized on the basis of residual dense network. Multiple skip-changed residual dense blocks are cascaded in feature extraction. Connection complexity is reduced to logarithm. The shallow feature in medical image is not processed repeatedly. The schema that meta-information embedding directly is used in feature map reconstruction. The universal knowledge between different scale factor task is learned through a small network. The diverse tasks are integrated. The tasks of different scale factors are transferred to the same dimension. The fractional scale factor is supported. The experimental results show that, compared with canonical method such as VDSR, the peak signal-to-noise ratio increases 0.17~1.57 dB and structural similarity increases 0.002 2~0.042 5.

Key words: super-resolution, deep learning, residual dense, meta-information embedding, medical image