新冠肺炎切割

Verion1.0~3.0 Author: ZhaoY

运行环境

Version	v3.0 20210605
编程语言	Python
Cuda版本	10.0
库	requirements

分割神经网络模型

网络名称	原始文章
$U^2$-Net	$U^2$-Net: Going Deeper with Nested U-Structure for Salient Object Detection

网络代码
U2-Net: Going Deeper with Nested U-Structure for Salient Object Detection
分割网络排行
Semantic Segmentation Semantic Segmentation on Cityscapes val

代码流程图

st=>start: start
rs=>operation: utils/rename_seg_pic
ps=>operation: utils/preprocessSeg
tr=>operation: segTrain
ts=>operation: segTest
vl=>operation: plot_heatmap and visualizeSeg
rc=>operation: utils/rename_clf_pic
pc=>operation: utils/preprocessClf
ic=>operation: infer or infer_multi_process
rd=>operation: Radiomics/exact_radiomics
e=>end: end

st->rs->ps->tr->ts->vl->rc->pc->ic->rd->e

1. 数据预处理

根据test_split_segmentation.txt与train_split_segmentation.txt将图片划分为训练集和测试集

从DATASET.md中可以知道数据掩膜的灰度值代表意义如下

mask中label表示如下(以png文件的灰度值区分)
	分类			    灰度值
	背景(BG)			  0
	肺野(LF)			  1
	肺磨玻璃影(GGO)      2
	肺实质(CO)		     3

我们分割的重点在于病灶，即肺磨玻璃影与肺实质，因此将背景和肺野归为一个总体的背景类

in_dir 数据目录，其有子目录imgs与masks

out_dir 输出npy数组目录，后续生成子目录imgs与masks+str(n_classes)

python ./utils/rename_seg_pic.py
python ./utils/preprocessSeg.py --in_dir data/seg/train/ --n_classes 3 --out_dir data/seg/process/train/
python ./utils/preprocessSeg.py --in_dir data/seg/train/ --n_classes 2 --out_dir data/seg/process/train/
python ./utils/preprocessSeg.py --in_dir data/seg/test/ --n_classes 3 --out_dir data/seg/process/test/
python ./utils/preprocessSeg.py --in_dir data/seg/test/ --n_classes 2 --out_dir data/seg/process/test/

注意！！！如果内存不够不要开z-score！！！分类网络有62000张图，开了z-score需要128G memory
but it can improve perfermance

2. 分割训练与测试

3-classes Train

python segTrain.py --model_name U2Net_n --num_classes 3 --normalize True --batch_size 4 --train_data_dir ./data/seg/process/train --val_data_dir ./data/seg/process/test --weight 1 20 20 --lrate 3e-4 --num_epochs 200

python segTrain_U2Net.py --model_name U2Net_n --num_classes 3 --normalize True --batch_size 4 --train_data_dir ./data/seg/process/train --val_data_dir ./data/seg/process/test --log_name U2Net_n0528-2050 --pth output/saved_models/U2Net_n/epoch_50_model.pth

3-classes Test

 python segTest.py --model_name U2Net_n --num_classes 3 --normalize True --test_data_dir ./data/seg/process/test --pth output_zscore/saved_models/U2Net_n/1_20_20_2/bestSegZcore.pth

2-classes Train

python segTrain.py --model_name U2Net_n_2c --num_classes 2 --normalize True --batch_size 4 --train_data_dir ./data/seg/process/train --val_data_dir ./data/seg/process/test --weight 1 20 --lrate 3e-4 --num_epochs 200

python segTrain.py --model_name U2Net_n_2c --num_classes 2 --normalize True --batch_size 4 --train_data_dir ./data/seg/process/train --val_data_dir ./data/seg/process/test --num_epochs 100 --weight 1 20 --lrate 3e-4 --pth output/saved_models/U2Net_n_2c/epoch_150_model.pth --log_name U2Net_n_2c0529-2000

2-classes Test

 python segTest.py --model_name U2Net_n --num_classes 2 --normalize True --test_data_dir ./data/seg/process/test --pth output_zscore/saved_models/U2Net_n/1_20_20_2/bestSegZcore.pth

热力图(test中调用不方便)

python plot_heatmap.py --model_name U2Net_n --out_dir output/segResult/ --num_classes 2
python plot_heatmap.py --model_name U2Net_n --out_dir output/segResult/ --num_classes 3

可视化

python visualizeSeg.py

导出requirements.txt

pipreqs . --encoding=utf8 --force

影像组学

python Radiomics/exact_radiomics.py

性能指标

Name	mAP	mPA	IoU	Mean Dice coff(mDC)	accurcy
Our(without z-score,3 classes)	0.6774	0.743	0.6346	0.7070	0.8932
Our(z-score,3 classes)	0.6906	0.7573	0.6514	0.7204	0.8941
Our(z-score,2 classes)	_	0.8842	_	_	_
UNet(cell)[1]	_	0.652	_	0.547	_
DRUNet(cell)[1]	_	0.658	_	0.562	_
FCN(cell)[1]	_	0.637	_	0.553	_
SegNet(cell)[1]	_	0.610	_	0.555	_
DeepLabv3(cell)[1]	_	0.662	_	0.587	_
Mask R-CNN(GGO+CO)[7]	0.5020	_	_	_	_
Mask R-CNN(Lesion)[7]	0.6192	_	_	_	_
Mask R-CNN*(z-score, Lesion)[7]	0.6602	_	_	_	_

3-class confusion matrix

z-score

0.998	0.200	0.077
0.002	0.747	0.154
0.0003	0.053	0.767

without z-score

0.998	0.218	0.096
0.002	0.710	0.128
0.0002	0.072	0.776

2-class confusion matrix (plot行列错误，但是数据是对的)

0.996	0.115
0.003	0.885

3. 分类

预处理分类图片

python ./utils/preprocessClf.py --in_dir /home/e201cv/Desktop/covid_data/clf/train /home/e201cv/Desktop/covid_data/clf/val /home/e201cv/Desktop/covid_data/clf/test --out_dir /home/e201cv/Desktop/covid_data/process_clf/train /home/e201cv/Desktop/covid_data/process_clf/val /home/e201cv/Desktop/covid_data/process_clf/test

获取掩模

python infer.py --infer_data_dirs /home/e201cv/Desktop/covid_data/process_clf/train /home/e201cv/Desktop/covid_data/process_clf/val /home/e201cv/Desktop/covid_data/process_clf/test --pth output_zscore/saved_models/U2Net_n/1_20_20_2/bestSegZcore.pth --num_classes 3 --device cuda

获取影像组学

python Radiomics/exact_radiomics.py --imgs_dir data/clf/train/imgs data/clf/val/imgs data/clf/test/imgs --masks_dir data/clf/train/masks data/clf/val/masks data/clf/test/masks --out_dir data/clf/train/radiomics data/clf/val/radiomics data/clf/test/radiomics

AnnLIU15/SegCovid