基于利尿肾图的新型肾实质自动分割算法

A Novel Automatic Renal Parenchyma Segmentation Algorithm Based on the Diuretic Renal Image

摘要: 尽管深度学习图像分割技术在医学图像处理中取得了良好的效果，但由于利尿肾动态显像具有噪声明显、对比度低、图像质量差和边界不清晰的特点，大多数现有的分割方法在获取感受野和提取图像特征信息方面仍然面临很大的挑战。为了解决上述问题，本文提出了一种新的肾实质自动分割算法，该算法采用的是Swin-Transformer编码器与解码器，并且结合了特征融合模块、连续扩张卷积模块与深度注意力模块。特征增强模块中的空间金字塔池模块，可以弥补空间表征并产生多尺度表征。连续扩张卷积能通过获取多尺度上下文聚合来扩大感受野。注意力模块通过顺序捕获多尺度编码器特征之间的关系来解决编码器和解码器特征之间的语义差距。该算法应用于私人的利尿肾造影数据集。仿真实验结果表明，与其他深度学习分割方法比较，该方法可以显著提高肾脏分割性能。

Abstract: Although deep learning image segmentation techniques have achieved good results in medical image processing, most of the existing segmentation methods still face great challenges in acquiring the sensory field and extracting the image feature information due to the fact that diuretic renal dynamic imaging is characterised by significant noise, low contrast, poor image quality, and unclear boundaries. In order to solve the above problems, this paper proposes a new automatic renal parenchyma segmentation algorithm, which employs a Swin-Transformer encoder and decoder, and combines a feature Enhancement module, a successive dilation convolution module, and a deep attention module. The spatial pyramid pooling module in the feature Enhancement module compensates for spatial representations and produces multi-scale representations. Sequential Expansion Convolution can expand the sensory field by capturing multiscale context aggregation. The Attention module addresses the semantic gap between encoder and decoder features by sequentially capturing relationships between multiscale encoder features. The algorithm is applied to a private diuretic nephrography dataset. The results of simulation experiments show that the method can significantly improve the performance of renal segmentation compared to other deep learning segmentation methods.