医学图像九分类
发布时间
阅读量：
阅读量
解决上一次九分类存在的问题
1.将所有图片名和图片的类别存在一个csv文件中，name对应图片的名字，label对应图片的标签。
def generate_csv(path,type,csv_path):
with open(csv_path,'w',newline='') as csvfile:
svwriter=csv.writer(csvfile,dialect="excel")
svwriter.writerow(['name', 'label'])
listdir=os.listdir(os.path.join(path,type))
for i in listdir:
classimage=os.path.join(path,type,i)#0,1,2,3,4
data=os.listdir(classimage)
for j in data:
j=j.split('.')[0]
svwriter.writerow([j,i])2.读取csv文件中对应的图片。由于csv文件中保存的是图片名称和图片的label，我们需要读取到具体的图片，此时我们需要重写dataset方法，将图片名和label放在字典中，由图片地址获得图片名，再根据图片名去字典中找对应图片的label，返回图片名以及label：
class mydataset(Dataset):
def __init__(self,data_folder,class_dict,transform=None):
self.data_folder=data_folder#存放图片的文件夹train,test,val
self.class_dict=class_dict #由csv得到的名，类别字典
self.transform=transform
self.imageclass=[s for s in os.listdir(data_folder)]
self.imagelist=[]
for i in self.imageclass:
self.path=os.path.join(data_folder,i)
self.ima=os.listdir(self.path)
for er in self.ima:
er=i+'/'+er
self.imagelist.append(er)
self.labels=[class_dict[i.split('.')[0].split('/')[1]] for i in self.imagelist]
def __len__(self):
return len(self.imagelist)
def __getitem__(self, idx):
image_path=os.path.join(self.data_folder,self.imagelist[idx])
img=Image.open(image_path)
imag=self.transform(img)
img_name=self.imagelist[idx].split('.')[0].split('/')[1]
label=self.class_dict[img_name]
return img_name,imag,labeldef init()方法用来初始化，我们必须实现def len()和def__getitem__()方法，getitem()方法每次返回一个图片，len()是整个的长度，需要调用getitem()多少次
由于class mydataset(Dataset)每次只能返回一个图片，因此我们需要使用dataloader来一次加载多张图片。
def data_prepare(path,type,batch_size,transform):
csv_data=pd.read_csv(path+'/'+type+'.csv')
class_dict={k:v for k,v in zip(csv_data.name,csv_data.label)}
#print(class_dict) # '名字'：类别
dataset=mydataset(data_folder=path+'/'+type,class_dict=class_dict,transform=transform)
dataloader=DataLoader(dataset=dataset,batch_size=batch_size,shuffle=True)
return dataloader使用resnet18进行迁移学习。
为了防止训练过拟合，引入了验证集。设置一个判断是否过拟合的条件，如果满足则不再训练，保存模型参数。进入测试阶段。
epoch=50
for i in range(epoch):
trainacc=train(model=model,Loss=loss,optimizer=optimizer,dataloader=train_loader)
valacc=val(model=model,Loss=loss,optimizer=optimizer,dataloader=val_loader)
if trainacc>0.9 and abs(trainacc-valacc)<=0.1:
bestacc=valacc
torch.save(model.state_dict(),'/home/cad429/code/yxy/Week2/codewek3/parameter.pkl')
break在训练和验证过程中为了计算每个类别对应的准确率，使用两个list，correct以及total，correct存放每个类别分类正确的数目，total每个类别的总数目。
使用pycm包来计算多分类的混淆矩阵。
完整代码：
import csv
import pandas as pd
import torch
import os
#将文件夹内的图片的名字存放在一个txt文件中。
import os
from pycm import *
from torchvision import transforms,models
import torch.nn as nn
from torch.autograd import Variable
from torch import optim
from torch.utils.data import Dataset,DataLoader
from PIL import Image
import math
from sklearn.metrics import confusion_matrix
from tensorboardX import SummaryWriter
path1=r"/home/cad429/code/yxy/Week2"
# path2=r"/home/cad429/code/yue/Week2/train2"
# path3=r"/home/cad429/code/yue/Week2/val"
csv_path="/home/cad429/code/yxy/Week2/test.csv"
csv_path2="/home/cad429/code/yxy/Week2/train2.csv"
csv_path3="/home/cad429/code/yxy/Week2/val.csv"
normalize = transforms.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225])
batch_size=4
transform = transforms.Compose([
transforms.Resize(size=(227, 227)),
transforms.RandomRotation(20),
transforms.RandomHorizontalFlip(),
transforms.ToTensor(),
normalize
])
bestacc=0
def generate_csv(path,type,csv_path):
with open(csv_path,'w',newline='') as csvfile:
svwriter=csv.writer(csvfile,dialect="excel")
svwriter.writerow(['name', 'label'])
listdir=os.listdir(os.path.join(path,type))
for i in listdir:
classimage=os.path.join(path,type,i)#0,1,2,3,4
data=os.listdir(classimage)
for j in data:
j=j.split('.')[0]
svwriter.writerow([j,i])
class mydataset(Dataset):
def __init__(self,data_folder,class_dict,transform=None):
self.data_folder=data_folder#存放图片的文件夹train,test,val
self.class_dict=class_dict #由csv得到的名，类别字典
self.transform=transform
self.imageclass=[s for s in os.listdir(data_folder)]
self.imagelist=[]
for i in self.imageclass:
self.path=os.path.join(data_folder,i)
self.ima=os.listdir(self.path)
for er in self.ima:
er=i+'/'+er
self.imagelist.append(er)
self.labels=[class_dict[i.split('.')[0].split('/')[1]] for i in self.imagelist]
def __len__(self):
return len(self.imagelist)
def __getitem__(self, idx):
image_path=os.path.join(self.data_folder,self.imagelist[idx])
img=Image.open(image_path)
imag=self.transform(img)
img_name=self.imagelist[idx].split('.')[0].split('/')[1]
label=self.class_dict[img_name]
return img_name,imag,label
def data_prepare(path,type,batch_size,transform):
csv_data=pd.read_csv(path+'/'+type+'.csv')
class_dict={k:v for k,v in zip(csv_data.name,csv_data.label)}
#print(class_dict) # '名字'：类别
dataset=mydataset(data_folder=path+'/'+type,class_dict=class_dict,transform=transform)
dataloader=DataLoader(dataset=dataset,batch_size=batch_size,shuffle=True)
return dataloader
generate_csv(path=path1,type='test',csv_path=csv_path)
generate_csv(path=path1,type='train2',csv_path=csv_path2)
generate_csv(path=path1,type='val',csv_path=csv_path3)
test_loader=data_prepare(path=path1,type='test',batch_size=batch_size,transform=transform)
train_loader=data_prepare(path=path1,type='train2',batch_size=batch_size,transform=transform)
val_loader=data_prepare(path=path1,type='val',batch_size=batch_size,transform=transform)
model=models.resnet18(pretrained=True)
for para in model.parameters():
para.require_grad=False
model.fc=nn.Linear(512,9)
optimizer=optim.SGD(model.parameters(),lr=0.001)
loss=nn.CrossEntropyLoss()
# model.eval()
# #model.cuda()
# for i,(img_name,img,label) in enumerate(test_loader):
# img=Variable(img)
# label=Variable(label)
# print("i 是：",i)
# print(label)
def train(model,Loss,optimizer,dataloader):
model.train()
model.cuda()
s=0
correcttrain = list(0 for i in range(9))
totaltrain = list(0 for i in range(9))
for i,(img_name,img,label) in enumerate(dataloader):
img=Variable(img).cuda()
label=Variable(label).cuda()
optimizer.zero_grad()
outputs=model(img)
loss=Loss(outputs,label)
loss.backward()
optimizer.step()
_,pred=torch.max(outputs,1)
pred=pred.cpu().numpy()
label=label.cpu().numpy()
for x in range(len(pred)):
a = pred[x]
if pred[x] == label[x]:
s+=1
correcttrain[a] += 1
totaltrain[a] += 1
else:
totaltrain[a] += 1
trainloader_len=len(dataloader)
acc=s/(trainloader_len*batch_size)
print("train acc:",acc)
return acc
def val(model,Loss,optimizer,dataloader):
model.cuda()
model.eval()
s1=0
correctval = list(0 for i in range(9))
totalval = list(0 for i in range(9))
for i,(ima_name,img,label) in enumerate(dataloader):
img=Variable(img).cuda()
label=Variable(label).cuda()
optimizer.zero_grad()
output=model(img)
loss = Loss(output, label)
_,pred=torch.max(output,1)
pred=pred.cpu().numpy()
label=label.cpu().numpy()
for j in range(len(pred)):
a = pred[j]
if pred[j] == label[j]:
s1+=1
correctval[a]+=1
totalval[a]+=1
else:
totalval[a]+=1
val_len=len(dataloader)
valacc=s1/(val_len*batch_size)
print("val acc :",valacc)
for i in range(9):
if totalval[i]!=0:
print(i,"类的准确率是：",correctval[i]/totalval[i])
return valacc
epoch=50
for i in range(epoch):
trainacc=train(model=model,Loss=loss,optimizer=optimizer,dataloader=train_loader)
valacc=val(model=model,Loss=loss,optimizer=optimizer,dataloader=val_loader)
if trainacc>0.9 and abs(trainacc-valacc)<=0.1:
bestacc=valacc
torch.save(model.state_dict(),'/home/cad429/code/yxy/Week2/codewek3/parameter.pkl')
break
#test
model.load_state_dict(torch.load('/home/cad429/code/yxy/Week2/codewek3/parameter.pkl'))
model.cuda()
model.eval()
ss=0
predtest=[]
labeltest=[]
for x,(img_name,img,label) in enumerate(test_loader):
img=Variable(img).cuda()
label=Variable(label).cuda()
output=model(img)
testloss=loss(output,label)
optimizer.zero_grad()
_,predd=torch.max(output,1)
predd=predd.cpu().numpy()
label=label.cpu().numpy()
for i in range(len(predd)):
if predd[i]==label[i]:
ss+=1
predtest.append(predd[i])
labeltest.append(label[i])
testlen=len(test_loader)
testacc=ss/(testlen*batch_size)
print("test acc is :",testacc)
cm=ConfusionMatrix(actual_vector=labeltest,predict_vector=predtest)
print(cm)测试的结果：
1.总的准确率：
测试集的混淆矩阵：
Predict 0 1 2 3 4 5 6 7 8
Actual
0 127 0 0 0 0 0 2 1 0
1 0 0 2 0 0 1 0 0 0
2 0 0 36 0 0 3 0 0 0
3 0 0 3 0 1 1 0 0 0
4 1 0 0 0 24 0 0 10 0
5 0 0 0 0 1 90 0 1 0
6 3 0 0 0 3 0 98 1 0
7 1 0 0 0 5 2 0 54 1
8 3 0 0 0 0 0 0 0 90
AUC(Area under the ROC curve) 0类 0.97927
1类 0.5
2类 0.95679
3类 0.5
4类 0.83342
5类0.98173
6类 0.96449
7类 0.91562
8类 0.98281
F1score：
F1(F1 score - harmonic mean of precision and sensitivity)
0类0.95849 1类 0.0
2类 0.9
3类 0.0
4类 0.69565
5类 0.95238
6类 0.9561
7类 0.83077
8类 0.97826
全部评论 (0)
还没有任何评论哟~