Target-aware U-Net with fuzzy skip connections for refined pancreas segmentation
Target-aware U-Net with fuzzy skip connections for refined pancreas segmentation
复制标题
DOI：
10.1016/j.asoc.2022.109818
复制
发表时间：
2022-11
期刊：
影响因子：
--
通讯作者：
Yufei Chen-;Chang Xu;Weiping Ding;Shicheng Sun;Xiaodong Yue;Hamido Fujita
中科院分区：
文献类型：
--
作者：
Yufei Chen-;Chang Xu;Weiping Ding;Shicheng Sun;Xiaodong Yue;Hamido Fujita
医学图像分割是胰腺疾病计算机辅助诊断的重要步骤之一。虽然已经提出了一些模型来处理胰腺的自动分割任务，但由于胰腺体积小，形状多变，边界不清晰，因此仍然具有挑战性。在本文中，我们提出了一个目标感知的U网（tU网），使用模糊跳跃连接胰腺分割。通过在U-Net结构中加入模糊跳跃连接模块，将低层特征转化为高层语义特征，有利于胰腺小而多变目标的分割。基于模糊特征映射，我们还设计了一个由全局平均池和深度卷积组成的目标注意机制。它通过增加重要通道的权重，使网络的解码器对目标特征更加敏感。所提出的方法进行评估的NIH数据集的82 CT卷，和胰腺医学分割十项全能（MSD）挑战数据集的281 CT卷。该模型取得了更好的结果与其他国家的最先进的模型相比。
Medical image segmentation is one of the important steps in the computer-aided diagnosis of pancreas diseases. Although some models have been proposed to deal with the task of automatic pancreas segmentation, it is still challenging due to the small size, variable shape and unclear boundary of pancreas. In this paper, we propose a target-aware U-Net (tU-Net) using fuzzy skip connection for pancreas segmentation. Through adding a fuzzy skip connection module into the U-Net architecture, the low-level features can be transformed into the high-level semantic features, which facilitates the segmentation of small and changeable targets of pancreas. Based on the fuzzy feature mapping, we also design a target attention mechanism consists of global average pooling and depthwise convolution. It makes the decoder of the network more sensitive to target features by increasing weights of important channels. The proposed method is evaluated on the NIH dataset of 82 CT volumes, and the pancreas Medical Segmentation Decathlon (MSD) challenge dataset of 281 CT volumes. The proposed model achieves better results comparing with other state-of-the-art models.