具有空间感知的3D U-Net模型
3D U-Net Model with Spatial Perception
冯鑫 1郝娟 1刘晓群1
作者信息
- 1. 河北建筑工程学院信息工程学院,河北 张家口 075000
- 折叠
摘要
针对脑肿瘤分割任务在输入多模态多通道的情况,提出了一种具有空间感知的焦点损失函数以及Soft Dice损失函数的3D U-Net模型.空间感知的加入为重要分割分区提供分割权重,提高重要分区的分割有效性.模型通过Dice系数和HD95(Hausdorff Distance 95th percentile)作为评估标准,在BraTS2021数据集上对比3D U-Net在肿瘤核心Dice系数提高3.21%,完整肿瘤Dice系数提高2.49%,增强肿瘤Dice系数提高3.1%,HD95系数减少4.90,显示出模型出色的分割能力.
Abstract
Aiming at the multi-modal and multi-channel input of brain tumor segmentation task,a 3D u-net model with spatial per-ception focus loss function and Soft Dice loss function is proposed.The addition of spatial awareness improves the segmentation weight for important partition and improves the segmentation effectiveness of important partition.The model uses Dice coefficient and HD95(Hausdorff Distance 95th percentile)as evaluation criteria.Compared with 3D U-Net on BraTS2021 dataset,the Dice co-efficient of tumor core increases by 3.21%,the Dice coefficient of complete tumor increases by 2.49%,the Dice coefficient of en-hanced tumor increases by 3.1%,and the HD95 coefficient decreases by 4.90,showing the excellent segmentation ability of the model.
关键词
脑肿瘤分割/3D/U-Net/空间感知Key words
brain tumor segmentation/3D U-Net/spatial perception引用本文复制引用
出版年
2024
NETL
NSTL
万方数据