一种神经母细胞瘤的PET/CT多模态图像分割方法

PET/CT Multimodal Image Segmentation Method for Neuroblastoma

目前关于PET/CT图像分割的方法缺少对图像全局上下文信息提取的考虑并且对提取到的特征信息进行融合的方式有待优化.针对上述问题,本文提出了并行提取和特征综合网络(Parallel Extraction and Feature Synthesis Network,PFNet).该网络首先利用卷积神经网络(convolutional neural network,CNN)和Transformer分别设计了两个支路分别从PET/CT图像中提取局部特征和全局上下文信息.其次,为了融合不同的特征信息,本文设计了一个特征综合模块,将两个分支提取的不同信息送入特征综合模块进行融合.在该模块中,设计了多策略融合单元,与此同时该模块利用通道注意力单元和依赖关系学习单元进一步学习通道特征信息和特征图自身依赖关系,以便进一步进行特征融合.此外本文引入了LogCoshDiceLoss作为损失函数,其利用函数Log-Cosh的平滑性可以提高分割性能.最后,在神经母细胞瘤数据集和HECKTOR 2021数据集上的实验结果表明本文提出的网络相比其它方法具有更好的分割性能,获得了较好的分割效果.

At present,the methods of PET/CT image segmentation lack the consideration of global context information extraction,and the fusion methods of the extracted feature information need to be optimized.To solve these problems,parallel extraction and feature synthesis network(named PFNet)is proposed.Firstly,the network uses convolutional neural network(CNN)and Transformer to design two branches to extract local features and global context information from PET/CT images respectively.Secondly,in order to fuse different feature information,this paper designs a feature synthesis module and sends the different information extracted by the two branches to the feature synthesis module for fusion.In this module,a multi-strategy fusion unit is designed.At the same time,the module uses the channel attention unit and the dependency learning unit to further learn the channel feature information and the dependency relationship of feature map for further feature fusion.In addition,LogCoshDiceLoss is used as the loss function,which can improve the segmentation performance by using the smoothness of function Log-Cosh.Finally,experimental results on the neuroblastoma dataset and the HECKTOR 2021 dataset show that the proposed network has better performance than other methods and achieves better segmentation results.