Automatic Brain Tumor Detection and Segmentation Using U-Net Based Fully Convolutional Networks_2017
2017-10-08 09:11
579 查看
作者:Hao Dong , Guang Yang 等
训练图像被插值为1*1*1,大小为240×240×155,然后经过数据标准化(每套多模态MRI减去自己的均值,除以标准差).
肿瘤被分了4类,1.necrosis(坏死), 2.edema(水肿), 3.non-enhancing, 4.enhancing tumor.使用FLAIR图像分割完整的肿瘤区域和除水肿之外的肿瘤,使用T1c(T1-weighted imaging with gadolinium enhancing contrast)被用来描述 enhancing tumor.
数据增强
如图:
flipping, rotation, shift and zoom比较简单,Shear能够轻微扭曲肿瘤水平方向的形状,但仍然不能获得足够的有变化的数据,因为肿瘤没有固定的形状.所以我们使用了elastic distortion,能够创造任意但是有合理形状的肿瘤.
U-Net Based Deep Convolutional Networks
网络结构如图一:
下采样层有5个卷积块,每块有两个卷积层,滤波器大小为3×3,步长为1.amxpooling步长为2×2.在上采样层,通过滤波器为3×3步长为2×2的deconvolutional层,将大小增大一倍,深度减小一半.每层大小不变.最后是一个1×1的卷积层,将深度降为2,分别反映前景和背景的分割.其他参数:
损失函数使用Soft Dice metric[25],如图:
优化器使用Adam(rate = 0.0001, maximum epochs = 100).通过正太分布(0均值,0.01方差)初始化所有变量,偏置为0.
评估:
TP,FP,FN分别代表真阳,假阳,假阴.
训练图像被插值为1*1*1,大小为240×240×155,然后经过数据标准化(每套多模态MRI减去自己的均值,除以标准差).
肿瘤被分了4类,1.necrosis(坏死), 2.edema(水肿), 3.non-enhancing, 4.enhancing tumor.使用FLAIR图像分割完整的肿瘤区域和除水肿之外的肿瘤,使用T1c(T1-weighted imaging with gadolinium enhancing contrast)被用来描述 enhancing tumor.
数据增强
如图:
flipping, rotation, shift and zoom比较简单,Shear能够轻微扭曲肿瘤水平方向的形状,但仍然不能获得足够的有变化的数据,因为肿瘤没有固定的形状.所以我们使用了elastic distortion,能够创造任意但是有合理形状的肿瘤.
U-Net Based Deep Convolutional Networks
网络结构如图一:
下采样层有5个卷积块,每块有两个卷积层,滤波器大小为3×3,步长为1.amxpooling步长为2×2.在上采样层,通过滤波器为3×3步长为2×2的deconvolutional层,将大小增大一倍,深度减小一半.每层大小不变.最后是一个1×1的卷积层,将深度降为2,分别反映前景和背景的分割.其他参数:
损失函数使用Soft Dice metric[25],如图:
优化器使用Adam(rate = 0.0001, maximum epochs = 100).通过正太分布(0均值,0.01方差)初始化所有变量,偏置为0.
评估:
TP,FP,FN分别代表真阳,假阳,假阴.
相关文章推荐
- Region-based Convolutional Networks for Accurate Object Detection and Segmentation----R-CNN论文笔记
- 【论文笔记】Region-based Convolutional Networks for Accurate Object Detection and Segmentation
- R-FCN: Object Detection via Region-based Fully Convolutional Networks
- Boundary-Aware Fully Convolutional Network for Brain Tumor Segmentation论文理解
- OverFeat: Integrated Recognition, Localization and Detection using Convolutional Networks
- OverFeat: Integrated Recognition, Localization and Detection using Convolutional Networks(阅读)
- R-FCN: Object Detection via Region-based Fully Convolutional Networks
- Large-scale Multimodal Gesture Segmentation and Recognition based on Convolutional Neural Networks
- 论文阅读 R-FCN: Object Detection via Region-based Fully Convolutional Networks
- OverFeat Integrated Recognition, Localization and Detection using Convolutional Networks
- 图像检测3-OverFeat-Integrated Recognition,Localization,and Detection using Convolutional Networks
- R-FCN: Object Detection via Region-based Fully Convolutional Networks 论文笔记
- 对 OverFeat: Integrated Recognition, Localization and Detection using Convolutional Networks 一文的理解
- Semantic Segmentation using Fully Convolutional Networks over the years
- 对论文R-FCN: Object Detection via Region-based Fully Convolutional Networks的阅读
- 论文阅读理解 - R-FCN: Object Detection via Region-based Fully Convolutional Networks
- R-FCN:Object Detection via Region-based Fully Convolutional Networks
- 论文笔记 | R-FCN: Object Detection via Region-based Fully Convolutional Networks
- 深度学习研究理解:OverFeat:Integrated Recognition, Localization and Detection using Convolutional Networks
- 深度学习研究理解6:OverFeat:Integrated Recognition, Localization and Detection using Convolutional Networks