项目GitHub主页：https://github.com/orobix/retina-unet
参考论文：Retina blood vessel segmentation with a convolution neural network (U-net)
1. 训练数据
1.1 训练图像、训练金标准随机分块
主代码：
# 训练集太少，采用分块的方法进行训练
def get_data_training(DRIVE_train_imgs_original, #训练图像路径
DRIVE_train_groudTruth, #金标准图像路径
patch_height,
patch_width,
N_subimgs,
inside_FOV):
train_imgs_original = load_hdf5(DRIVE_train_imgs_original)
train_masks = load_hdf5(DRIVE_train_groudTruth)
#visualize(group_images(train_imgs_original[0:20,:,:,:],5),'imgs_train').show()
train_imgs = my_PreProc(train_imgs_original) # 图像预处理 归一化等
train_masks = train_masks/255.
train_imgs = train_imgs[:,:,9:574,:] # 图像裁剪 size=565*565
train_masks = train_masks[:,:,9:574,:] # 图像裁剪 size=565*565
data_consistency_check(train_imgs,train_masks) # 训练图像和金标准图像一致性检查
assert(np.min(train_masks)==0 and np.max(train_masks)==1) #金标准图像 2类 0-1
print ("\n train images/masks shape:")
print (train_imgs.shape)
print ("train images range (min-max): " +str(np.min(train_imgs)) +' - '+str(np.max(train_imgs)))
print ("train masks are within 0-1\n")
# 从整张图像中-随机提取-训练子块
patches_imgs_train, patches_masks_train =
extract_random(train_imgs,train_masks,patch_height,patch_width,N_subimgs,inside_FOV)
data_consistency_check(patches_imgs_train, patches_masks_train) # 训练图像子块和金标准图像子块一致性检查
print ("\n train PATCHES images/masks shape:")
print (patches_imgs_train.shape)
print ("train PATCHES images range (min-max): " +
str(np.min(patches_imgs_train)) +' - '+str(np.max(patches_imgs_train)))
return patches_imgs_train, patches_masks_train随机提取子块：
# 训练集图像 随机 提取子块
def extract_random(full_imgs,full_masks, patch_h,patch_w, N_patches, inside=True):
if (N_patches%full_imgs.shape[0] != 0): # 检验每张图像应该提取多少块
print "N_patches: plase enter a multiple of 20"
exit()
assert (len(full_imgs.shape)==4 and len(full_masks.shape)==4) # 张量尺寸检验
assert (full_imgs.shape[1]==1 or full_imgs.shape[1]==3) # 通道检验
assert (full_masks.shape[1]==1) # 通道检验
assert (full_imgs.shape[2] == full_masks.shape[2] and full_imgs.shape[3] == full_masks.shape[3]) # 尺寸检验
patches = np.empty((N_patches,full_imgs.shape[1],patch_h,patch_w)) # 训练图像总子块
patches_masks = np.empty((N_patches,full_masks.shape[1],patch_h,patch_w)) # 训练金标准总子块
img_h = full_imgs.shape[2]
img_w = full_imgs.shape[3]
patch_per_img = int(N_patches/full_imgs.shape[0]) # 每张图像中提取的子块数量
print ("patches per full image: " +str(patch_per_img))
iter_tot = 0 # 图像子块总量计数器
for i in range(full_imgs.shape[0]): # 遍历每一张图像
k=0 # 每张图像子块计数器
while k <patch_per_img:
x_center = random.randint(0+int(patch_w/2),img_w-int(patch_w/2)) # 块中心的范围
y_center = random.randint(0+int(patch_h/2),img_h-int(patch_h/2))
if inside==True:
if is_patch_inside_FOV(x_center,y_center,img_w,img_h,patch_h)==False:
continue
patch = full_imgs[i,:,y_center-int(patch_h/2):y_center+int(patch_h/2),
x_center-int(patch_w/2):x_center+int(patch_w/2)]
patch_mask = full_masks[i,:,y_center-int(patch_h/2):y_center+int(patch_h/2),
x_center-int(patch_w/2):x_center+int(patch_w/2)]
patches[iter_tot]=patch # size=[Npatches, 3, patch_h, patch_w]
patches_masks[iter_tot]=patch_mask # size=[Npatches, 1, patch_h, patch_w]
iter_tot +=1 # 子块总量计数器
k+=1 # 每张图像子块总量计数器
return patches, patches_masks数据一致性检查函数：
# 训练集图像 和 金标准图像一致性检验
def data_consistency_check(imgs,masks):
assert(len(imgs.shape)==len(masks.shape))
assert(imgs.shape[0]==masks.shape[0])
assert(imgs.shape[2]==masks.shape[2])
assert(imgs.shape[3]==masks.shape[3])
assert(masks.shape[1]==1)
assert(imgs.shape[1]==1 or imgs.shape[1]==3)1.2 训练金标准改写成Une输出形式
# 将金标准图像改写成模型输出形式
def masks_Unet(masks): # size=[Npatches, 1, patch_height, patch_width]
assert (len(masks.shape)==4)
assert (masks.shape[1]==1 )
im_h = masks.shape[2]
im_w = masks.shape[3]
masks = np.reshape(masks,(masks.shape[0],im_h*im_w)) # 单像素建模
new_masks = np.empty((masks.shape[0],im_h*im_w,2)) # 二分类输出
for i in range(masks.shape[0]):
for j in range(im_h*im_w):
if masks[i,j] == 0:
new_masks[i,j,0]=1 # 金标准图像的反转
new_masks[i,j,1]=0 # 金标准图像
else:
new_masks[i,j,0]=0
new_masks[i,j,1]=1
return new_masks2. 网络输出转换成图像子块
# 网络输出 size=[Npatches, patch_height*patch_width, 2]
def pred_to_imgs(pred, patch_height, patch_width, mode="original"):
assert (len(pred.shape)==3)
assert (pred.shape[2]==2 ) # 确认是否为二分类
pred_images = np.empty((pred.shape[0],pred.shape[1])) #(Npatches,height*width)
if mode=="original": # 网络概率输出
for i in range(pred.shape[0]):
for pix in range(pred.shape[1]):
pred_images[i,pix]=pred[i,pix,1] #pred[:, :, 0] 是反分割图像输出 pred[:, :, 1]是分割输出
elif mode=="threshold": # 网络概率-阈值输出
for i in range(pred.shape[0]):
for pix in range(pred.shape[1]):
if pred[i,pix,1]>=0.5:
pred_images[i,pix]=1
else:
pred_images[i,pix]=0
else:
print ("mode " +str(mode) +" not recognized, it can be 'original' or 'threshold'")
exit()
# 改写成(Npatches,1, height, width)
pred_images = np.reshape(pred_images,(pred_images.shape[0],1, patch_height, patch_width))
return pred_images3. 测试图像按顺序分块、预测子块重新整合成图像
3.1 测试图像分块
def get_data_testing_overlap(DRIVE_test_imgs_original,
DRIVE_test_groudTruth,
Imgs_to_test, # 20
patch_height,
patch_width,
stride_height,
stride_width):
test_imgs_original = load_hdf5(DRIVE_test_imgs_original)
test_masks = load_hdf5(DRIVE_test_groudTruth)
test_imgs = my_PreProc(test_imgs_original)
test_masks = test_masks/255.
test_imgs = test_imgs[0:Imgs_to_test,:,:,:]
test_masks = test_masks[0:Imgs_to_test,:,:,:]
test_imgs = paint_border_overlap(test_imgs, patch_height, # 拓展图像 可以准确划分
patch_width, stride_height, stride_width)
assert(np.max(test_masks)==1 and np.min(test_masks)==0)
print ("\n test images shape:")
print (test_imgs.shape)
print ("\n test mask shape:")
print (test_masks.shape)
print ("test images range (min-max): " +str(np.min(test_imgs)) +' - '+str(np.max(test_imgs)))
# 按照顺序提取图像快 方便后续进行图像恢复（作者采用了overlap策略）
patches_imgs_test = extract_ordered_overlap(test_imgs,patch_height,patch_width,stride_height,stride_width)
print ("\n test PATCHES images shape:")
print (patches_imgs_test.shape)
print ("test PATCHES images range (min-max): " +
str(np.min(patches_imgs_test)) +' - '+str(np.max(patches_imgs_test)))
return patches_imgs_test, test_imgs.shape[2], test_imgs.shape[3], test_masks #原始大小原始图像进行拓展填充：
def paint_border_overlap(full_imgs, patch_h, patch_w, stride_h, stride_w):
assert (len(full_imgs.shape)==4) #4D arrays
assert (full_imgs.shape[1]==1 or full_imgs.shape[1]==3) #check the channel is 1 or 3
img_h = full_imgs.shape[2] #height of the full image
img_w = full_imgs.shape[3] #width of the full image
leftover_h = (img_h-patch_h)%stride_h #leftover on the h dim
leftover_w = (img_w-patch_w)%stride_w #leftover on the w dim
if (leftover_h != 0): #change dimension of img_h
tmp_full_imgs = np.zeros((full_imgs.shape[0],full_imgs.shape[1],img_h+(stride_h-leftover_h),img_w))
tmp_full_imgs[0:full_imgs.shape[0],0:full_imgs.shape[1],0:img_h,0:img_w] = full_imgs
full_imgs = tmp_full_imgs
if (leftover_w != 0): #change dimension of img_w
tmp_full_imgs = np.zeros((full_imgs.shape[0],full_imgs.shape[1],full_imgs.shape[2],img_w+(stride_w - leftover_w)))
tmp_full_imgs[0:full_imgs.shape[0],0:full_imgs.shape[1],0:full_imgs.shape[2],0:img_w] = full_imgs
full_imgs = tmp_full_imgs
return full_imgs
按顺序提取图像子块：
# 按照顺序对拓展后的图像进行子块采样
def extract_ordered_overlap(full_imgs, patch_h, patch_w,stride_h,stride_w):
assert (len(full_imgs.shape)==4)
assert (full_imgs.shape[1]==1 or full_imgs.shape[1]==3)
img_h = full_imgs.shape[2]
img_w = full_imgs.shape[3]
assert ((img_h-patch_h)%stride_h==0 and (img_w-patch_w)%stride_w==0)
N_patches_img = ((img_h-patch_h)//stride_h+1)*((img_w-patch_w)//stride_w+1) # 每张图像采集到的子图像
N_patches_tot = N_patches_img*full_imgs.shape[0] # 测试集总共的子图像数量
patches = np.empty((N_patches_tot,full_imgs.shape[1],patch_h,patch_w))
iter_tot = 0
for i in range(full_imgs.shape[0]):
for h in range((img_h-patch_h)//stride_h+1):
for w in range((img_w-patch_w)//stride_w+1):
patch = full_imgs[i,:,h*stride_h:(h*stride_h)+patch_h,w*stride_w:(w*stride_w)+patch_w]
patches[iter_tot]=patch
iter_tot +=1 #total
assert (iter_tot==N_patches_tot)
return patches 3.2 对于图像子块进行复原
# [Npatches, 1, patch_h, patch_w] img_h=new_height[588] img_w=new_width[568] stride-[10,10]
def recompone_overlap(preds, img_h, img_w, stride_h, stride_w):
assert (len(preds.shape)==4) # 检查张量尺寸
assert (preds.shape[1]==1 or preds.shape[1]==3)
patch_h = preds.shape[2]
patch_w = preds.shape[3]
N_patches_h = (img_h-patch_h)//stride_h+1 # img_h方向包括的patch_h数量
N_patches_w = (img_w-patch_w)//stride_w+1 # img_w方向包括的patch_w数量
N_patches_img = N_patches_h * N_patches_w # 每张图像包含的patch的数目
assert (preds.shape[0]%N_patches_img==0
N_full_imgs = preds.shape[0]//N_patches_img # 全幅图像的数目
full_prob = np.zeros((N_full_imgs,preds.shape[1],img_h,img_w))
full_sum = np.zeros((N_full_imgs,preds.shape[1],img_h,img_w))
k = 0 #迭代所有的子块
for i in range(N_full_imgs):
for h in range((img_h-patch_h)//stride_h+1):
for w in range((img_w-patch_w)//stride_w+1):
full_prob[i,:,h*stride_h:(h*stride_h)+patch_h,w*stride_w:(w*stride_w)+patch_w]+=preds[k]
full_sum[i,:,h*stride_h:(h*stride_h)+patch_h,w*stride_w:(w*stride_w)+patch_w]+=1
k+=1
assert(k==preds.shape[0])
assert(np.min(full_sum)>=1.0)
final_avg = full_prob/full_sum # 叠加概率 / 叠加权重 ： 采用了均值的方法
print final_avg.shape
assert(np.max(final_avg)<=1.0)
assert(np.min(final_avg)>=0.0)
return final_avg