基于多尺度细节增强的脑瘤图像分割研究
Brain tumor image segmentation based on multi-scale detail enhancement
刘遵雄 1陈子涵 1蔡体健 1陈均 1罗词勇1
作者信息
- 1. 华东交通大学信息工程学院,江西南昌 330013
- 折叠
摘要
针对脑瘤图像分割网络的跳跃连接引起的语义特征传递不平衡,以及多尺度特征关联不足,导致细节特征丢失,从而造成对细小目标肿瘤的分割精度不佳的问题,提出一种改进的Res-Unet框架的分割模型.该模型引入多尺度注意力融合模块,通过混合多尺度特征使模型更好地适应不同尺寸的肿瘤;该模型在跳跃连接中引入空间注意力模块,增强特征表达同时避免无用信息的干扰,保留特征图空间细节特征;最后通过辅助分类器模块,在解码器部分对不同尺度特征图进行特征预测.使用BraTS2020数据集进行实验和评估,使用Dice系数来评估模型分割效果.结果显示,改进的网络在全肿瘤区域、肿瘤核心区域和增强肿瘤区域的Dice系数分别为0.8877、0.8229、0.8027,相比于通道注意力模型增强肿瘤区域和肿瘤核心区域的系数分别提升2.6%和0.14%,证明改进模型在脑瘤核磁共振图像分割的有效性和精确性.
Abstract
Given that the imbalance of semantic feature transfer caused by the skip connection of the brain tumor image segmentation network and the insufficient correlation of multi-scale features lead to the loss of details,resulting in poor segmentation accuracy for small tumors,an improved segmentation model of the Res-Unet framework is proposed.The model introduces a multi-scale attention fusion module which makes the model better adaptable to tumors of different sizes by mixing multi-scale features,and adds a spatial attention module to the skip connection to enhance feature expression while avoiding the interference of useless information,preserving the spatial details of feature maps.Through the auxiliary classifier module,the decoder performs feature prediction on feature maps of different scales.The BraTS2020 dataset is used for experiments and evaluations,and the model segmentation performance is evaluated with Dice score.The results show that the improved network achieved average Dice scores of 0.887 7,0.822 9,and 0.802 7 for whole tumor,tumor core,and enhancing tumor,respectively.Compared with the channel attention model,the improved model increases the scores of enhancing tumor and tumor core by 2.6%and 0.14%,respectively,which proves its effectiveness and accuracy for brain tumor MR image segmentation.
关键词
脑肿瘤/图像分割/注意力机制/辅助分类器Key words
brain tumor/image segmentation/attention mechanism/auxiliary classifier引用本文复制引用
基金项目
国家自然科学基金(62166018)
江西省重点研发计划(20203BBE53029)
江 西 省 自 然 科 学 基 金(20232BAB202055)
出版年
2024
NETL
NSTL
万方数据