基于不确定性引导和尺度一致性的肾肿瘤图像分割方法  

Kidney Tumor Image Segmentation Method Based on Uncertainty Guidance and Scale Consistency

作　　者：侯冰震 张桂梅[1] 彭昆 HOU Bingzhen;ZHANG Guimei;PENG Kun(Institute of Computer Vision,Nanchang Hangkong University,Nanchang 330063)

机构地区：[1]南昌航空大学计算机视觉研究所,南昌330063

出　　处：《模式识别与人工智能》2023年第2期95-107,共13页Pattern Recognition and Artificial Intelligence

基　　金：国家自然科学基金项目(No.62261039)资助。

摘　　要：针对肾肿瘤图像分割中的边界模糊和特征丢失问题,在RAUNet(Residual Attention U-Net)的基础上,提出基于不确定性引导和尺度一致性的肾肿瘤图像分割方法.针对肾肿瘤图像边界模糊问题,在解码层引入不确定性引导模块,根据不确定性自适应分配权重,弱化错误像素点的影响,提高模型的边界定位能力.针对下采样引起的特征丢失问题,提出尺度注意力模块和特征一致性损失,利用多尺度融合策略融合不同尺度特征,并与编码器底部特征进行尺度一致性约束,缓解特征丢失问题.在公开数据集KiTS19上的肾脏和肾肿瘤的图像分割实验表明,文中方法提高肾肿瘤的分割精度.此外,由于文中方法引入不确定性引导模块,分割结果具有较好的可靠性.Aiming at the problems of boundary blurring and feature loss in kidney tumor image segmentation,a kidney tumor image segmentation method based on uncertainty guidance and scale consistency is proposed on the basis of residual attention U-net model.For blurred boundaries of kidney tumor images,an uncertainty guidance module is introduced into the decoding layer to allocate weights adaptively based on uncertainty.Thus,the effect of wrong pixels is reduced and the boundary localization ability of the model is improved.For the problem of feature loss caused by down-sampling,the scale attention module and feature consistency loss are proposed.The multi-scale fusion strategy is utilized to fuse features of different scales,and the scale consistency constraint is conducted with the features at the bottom of the encoder to alleviate the problem of feature loss.Finally,experiments of kidney and kidney tumor segmentation on the public dataset KiTS19 demonstrate that the proposed segmentation method greatly improves the segmentation accuracy.In addition,the segmentation results of the proposed method hold better reliability due to the uncertainty guidance module.

关 键 词：图像分割 不确定性引导 尺度一致性 注意力模块 肾肿瘤 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]