基于多尺度特征和反向注意力的肝脏肿瘤自动分割方法

Automatic liver tumor segmentation using multi-scale features and reverse attention

肝脏肿瘤分割旨在定位肝脏肿瘤区域,以辅助医生进行精准诊治。鉴于深度学习能自动学习医学图像中复杂的特征和结构,已成为肝脏肿瘤分割的主流方法之一。但肝脏肿瘤的大小、形态存在显著差异及边缘模糊等问题,限制了深度学习模型的分割性能。基于此,该文提出了一种融合多尺度特征和反向注意力机制的深度网络,并用于肝脏肿瘤的自动分割。具体地,基于U-Net模型的框架,分别设计了多尺度特征提取模块和基于深度监督的反向注意力模块,使得网络能根据分割目标的大小自适应地选择不同尺度的特征,并引导网络关注分割目标的边缘特征,进而提高网络的边缘分割能力。此外,设计了一种新的混合损失,以解决医学图像分割中的类别不平衡问题。最后,在MICCAI2017 LiTS挑战赛数据集的数值实验结果表明,所提方法的Dice系数、平均对称表面距离ASSD分别为76.12%和3.25 mm。

The objective of liver tumor segmentation is to pinpoint the region containing the liver tumor,aiding medical professionals in delivering precise diagnosis and treatment.Owing to its capacity to automatically learn intricate features and structures from medical images,deep learning has emerged as one of the mainstream approaches to liver tumor segmentation.Nonetheless,the size,shape,and blurred edges of liver tumor exhibit significant variations,which constrain the segmentation performance of deep learning models.Consequently,we introduce a deep network that incorporates multi-scale features and a reverse attention mechanism for the automatic segmentation of liver tumor.Specifically,drawing upon the U-Net model framework,this study designs a multi-scale feature extraction module and a reverse attention module based on deep supervision.This enables the network to adaptively select features of varying scales according to the size of the segmented target and direct the network’s attention towards the edge features of the segmented object,thereby enhancing the network’s edge segmentation capability.Furthermore,a novel hybrid loss is devised to address the issue of class imbalance in medical image segmentation.The recent numerical experiments conducted on the MICCAI2017 LiTS Challenge dataset have demonstrated that the proposed method achieves a Dice coefficient of 76.12%and an average symmetric surface distance(ASSD)of 3.25 mm.