SDRNet:基于空间信息恢复的医学图像分割网络

SDRNet: spatial detail recovery network for medical image segmentation

利用深度神经网络解决医学图像分割任务的模型有很多,但这些模型普遍忽视了深度神经网络中由多次下采样操作导致的空间细节丢失的问题,而空间细节中包含大量小区块,边缘等关键信息,这些细节的丢失会导致模型整体性能下降,也会使得分割结果的医疗辅助价值降低。本文提出了一种空间信息恢复网络(SDRNet)来解决上述难题。首先本文提出了空间信息注意力分支(SDAB)优化空间信息的抽取与表达,同时降低对语义信息的干扰;然后提出了特征强化模块(FEM)增强模型对语义信息的编码表达能力,优化训练过程。LUNA数据集上的实验结果表明,提出的模块能协同工作,更好地处理边缘细节和小区块,SDRNet在2个模块的协同作用下能实现更优的分割性能,超越了对比的经典方法,实现了96.44%的平均交并比。

Many models that use deep neural networks have been used to perform medical image segmentation tasks.However, these models generally ignore the problem of loss of spatial details caused by multiple downsampling operations in deep neural networks, and usually the spatial details contain a large number of small blocks and edges. The loss of these spatial details may cause overall performance decline of the model, and reduce the auxiliary medical value of the segmentation result. In this paper, a spatial detail recovery network(SDRNet) is proposed to solve the above problems. First, the spatial detail attention branch(SDAB) is proposed to optimize the extraction and representation of spatial information, which can also reduce the interference to semantic information. Then the feature enhancement module(FEM) is proposed to enhance the model’s ability to express semantic information, and optimize the training process. Experimental results on the LUNA dataset show that the proposed modules can work together and better segment edge details and small blocks. SDRNet can achieve better segmentation performance with synergy of the two modules, outperforming the classic methods of comparison, and achieving 96.44% mean IOU.