Brain Tumors

项目名称:脑肿瘤分割

编程语言:Keras/Python


环境配置链接: 3D U-Net

1. Interface N4BiasFieldCorrection failed to run.
I am really new in this area. When I import  N4BiasFieldCorrection from nipype.interfaces.ants, it works well. But when I run the python code, the problem shows : OSError: command ‘N4BiasFieldCorrection’ could not be found on host chenruideMacBook-Pro.local Interface N4BiasFieldCorrection failed to run. Anything else I should import or download ?
      Nipype only gets you the interface to third-party software such as FSL and ANTS. You do not get the software when you install nipype. You can try to run N4BiasFieldCorrection on your prompt and it will give you the same message: command not found.
Please install ANTS in your local machine!!!

2. Process finished with exit code 139(interruped by signal 11:SIGSEGV)
电脑配置 nvidia 1060
错误原因:电脑配置太low,数据存储格式是h5格式的,训练到epoch的最后一次获取验证数据的时候,读取h5数据太慢,而训练的速度太快导致获取不到数据,即验证数组为空,出现段错误。
解决办法:
a. 使用npz存储格式代替h5存储格式,但是这样的缺点是压缩率变低,数据包变大,但是可以训练
b. 可以不修改h5,在进入get_training_and_validation_generator这个函数后,把h5中的数据直接提取出来换成数组即可

3. N4ITKBiasFieldCorrection介绍
paper:Intensity non-uniformity correction using N3 on 3-T scanners with multichannel phased array coils


Process:


Model:

paper:Multi-level Activation for Segmentation of Hierarchically-nested Classes on 3D-Unet


Tricks:

MethodsName
NormalizationBN / LN / IN / SN
DroupoutS dropout / dropblock
Reduce dim1*1 kernel
DilatedHDC/ASPP
Up samplingdeconvolution / bilinear / DUC
Crfdense crf
Early stopyes
Lossdice loss / jaccard loss / focal loss
Regularizationl1_l2 / W constraint
ActivationReLU / PreLU / LeakyReLU

结果:

配置:显卡1060, 图片使用112*112*112
训练:Epoch 175/500 loss: -0.7303 – jaccard: 0.6066 – val_loss: -0.6615 – val_jaccard: 0.5421
Scores:
1.WT:0.8886     
2.TC:0.8049     
3.ET:0.7772

损失函数和箱状图:

Dice Loss
Box Diagram

分割结果:

3D 效果
左边是ground truth, 右边是预测结果

本文总结于网络文章,加入了个人理解,仅用于个人学习研究,不得用于其他用途,如涉及版权问题,请联系邮箱513403849@qq.com

Leave a Reply

Your email address will not be published. Required fields are marked *