Scale-adaptive super-feature based MetricUNet for brain tumor segmentation
Scale-adaptive super-feature based MetricUNet for brain tumor segmentation
复制标题
DOI：
10.1016/j.bspc.2021.103442
复制
发表时间：
2022-03
期刊：
影响因子：
--
通讯作者：
Yujian Liu-;Jie Du;C. Vong;Guanghui Yue;Juan Yu;Yuli Wang;Baiying Lei;Tianfu Wang
中科院分区：
文献类型：
--
作者：
Yujian Liu-;Jie Du;C. Vong;Guanghui Yue;Juan Yu;Yuli Wang;Baiying Lei;Tianfu Wang
脑肿瘤的准确分割对脑肿瘤的诊断和治疗方案至关重要。一般来说，脑肿瘤包括WT（全瘤）、TC（肿瘤核心）和ET（增强肿瘤），TC和ET在临床上比WT重要得多。然而，TC和ET通常包含模糊的边界，占用的像素比WT少得多。最近，提出了基于体素-度量学习的MetricUNet，该方法考虑图像中体素级的特征关系，以获得更精细的分割结果。然而，它可能不适用于脑肿瘤的分割。这是因为脑肿瘤的尺度/大小在图像之间差异很大，导致MetricUNet的模型训练无效。此外，在脑肿瘤分割中考虑体素级特征关系的计算量较大。在这项工作中，提出了基于尺度自适应超特征的MetricUNet(S2MetricUNet)，并提供了两个优点：i)由于提出了一种新颖的尺度自适应度量损失，可以在解决图像之间的尺度变化的同时学习更多关于TC和ET的上下文信息，因此在TC和ET上具有更高的精度；Ii)大大减少了计算量，因为提出了超体素级特征来表示非边缘区域中的一组体素级特征（具有相同标签）。在公共数据集BraTS2019上的实验结果表明，就Dice而言，我们的方法在TC上的改进高达3.38%，在ET上的改进高达3.82%。此外，我们的S2MetricUNet的计算量减少到MetricUNet的1/11左右。
Accurate segmentation of brain tumors is very essential for brain tumor diagnosis and treatment plans. In general, brain tumor includes WT (whole tumor), TC (tumor core) and ET (enhance tumor), and TC and ET are much more important than WT clinically. However, TC and ET usually contain blurred boundaries, and occupy much fewer pixels than WT. Recently, MetricUNet based on voxel-metric learning is proposed, which considers voxel-level feature relationship in the image to obtain finer segmentation results. However, it may not be applicable in brain tumor segmentation. That is because the scales/sizes of brain tumor greatly vary between images and causing ineffective model training in MetricUNet. Moreover, it has heavy computation for considering voxel-level feature relationship in brain tumor segmentation. In this work, aScale-adaptive Super-feature based MetricUNet(S2MetricUNet) is proposed and provides two advantages: i) higher accuracy on TC and ET since a novelscale-adaptive metric lossis proposed for learning of more context information about TC and ET while addressing the scale variation between images; ii) significant reduction on computation since asuper voxel-level featureis proposed to represent a group of voxel-level features (of the same label) in non-edge regions. The experimental results on public dataset BraTS2019 have demonstrated that the improvement of our method is up to 3.38% on TC and 3.82% on ET in terms Dice. Moreover, the computation of our S2MetricUNet is reduced to about 1/11 of MetricUNet.