基于回环残差注意力机制U-net的胰腺分割

器官自动分割是医学图像分析中的一个重要而具有挑战性的问题.胰腺是一种位于腹部内部的软组织器官,其欠缺有形器官有固定形状的特点.由于胰腺周围的重要结构组织的关系紧密且多变,且有边缘界限不易确定等特点,采用传统的分割方法存在受噪声影响大、过分割和欠分割等问题,难以达到很高的准确率.虽然基于深度卷积神经网络的U-net很好地解决了传统分割方法中的一系列问题,但仍然存在器官形状分割不清晰,以及存在较多毛刺的问题.提出一种基于回环残差注意力机制U-net(ringed residual attention U-net,RRA U-net)的胰腺分割方法,通过加入注意力机制增大有效特征的占重比以及引入回环残差结构来增强鉴别特征的能力和加快网络的收敛速度,最终达到了较高的分割准确率.

Pancreas segmentation based on ringed residual attention U-net

Organ segmentation is an important and challenging task in medical image analysis. The pancreas is a soft tissue organ located inside the abdomen, and the lack of tangible organs is characterized by a fixed shape. The important structural tissues around the pancreas are closely related and changeable, and the edge boundaries are not easy to determine. The traditional segmentation methods have the problems of being affected by noise, over-segmentation and under-segmentation. Although the U-net network based on deep convolutional neural networks solves a series of problems in traditional segmentation methods, there are still problems with unclear segmentation of organ shapes and more burrs. Therefore, this paper proposes a new network, named ringed residual attention U-net (RRA U-net) for pancreas segmentation, which increases the proportion of effective features by adding an attention mechanism and introduces the loop residual structure to enhance the ability to identify features and accelerate the convergence rate of the network, and achieves a high segmentation accuracy.