牙齿的X射线图像的语义分割（附代码）

第一步结果

第二步结果

 

本研究的目的是通过使用U-Net模型和二值图像分析的深度学习方法，在一次拍摄的全景x射线图像中自动语义分割和测量牙齿的总长度，以便为牙齿疾病、疾病和状况的管理提供诊断信息。

模型结构：

代码：

# -*- coding: utf-8 -*-



#### MODEL ###
from tensorflow.keras.layers import Input, Conv2D, MaxPooling2D, BatchNormalization,concatenate,Conv2DTranspose,Dropout
from tensorflow.keras.models import Model


def UNET (input_shape=(512,512,1),last_activation='sigmoid'):
    inputs=Input(shape=input_shape)
    
    conv1 = Conv2D(32,(3,3), activation = 'relu', padding = 'same', kernel_initializer = 'he_normal')(inputs)
    d1=Dropout(0.1)(conv1)
    conv2 = Conv2D(32,(3,3), activation = 'relu', padding = 'same', kernel_initializer = 'he_normal')(d1)
    b=BatchNormalization()(conv2)
    
    pool1 = MaxPooling2D(pool_size=(2, 2))(b)
    conv3 = Conv2D(64,(3,3), activation = 'relu', padding = 'same', kernel_initializer = 'he_normal')(pool1)
    d2=Dropout(0.2)(conv3)
    conv4 = Conv2D(64,(3,3), activation = 'relu', padding = 'same', kernel_initializer = 'he_normal')(d2)
    b1=BatchNormalization()(conv4)
    
    pool2 = MaxPooling2D(pool_size=(2, 2))(b1)
    conv5 = Conv2D(128,(3,3), activation = 'relu', padding = 'same', kernel_initializer = 'he_normal')(pool2)
    d3=Dropout(0.3)(conv5)
    conv6 = Conv2D(128,(3,3), activation = 'relu', padding = 'same', kernel_initializer = 'he_normal')(d3)
    b2=BatchNormalization()(conv6)
    
    pool3 = MaxPooling2D(pool_size=(2, 2))(b2)
    conv7 = Conv2D(256,(3,3), activation = 'relu', padding = 'same', kernel_initializer = 'he_normal')(pool3)
    d4=Dropout(0.4)(conv7)
    conv8 = Conv2D(256,(3,3), activation = 'relu', padding = 'same', kernel_initializer = 'he_normal')(d4)
    b3=BatchNormalization()(conv8)
    
    pool4 = MaxPooling2D(pool_size=(2, 2))(b3)
    conv9 = Conv2D(512,(3,3),activation = 'relu', padding = 'same', kernel_initializer = 'he_normal')(pool4)
    d5=Dropout(0.5)(conv9)
    conv10 = Conv2D(512,(3,3), activation = 'relu', padding = 'same', kernel_initializer = 'he_normal')(d5)
    b4=BatchNormalization()(conv10)
    
    
    conv11 = Conv2DTranspose(512,(4,4), activation = 'relu', padding = 'same', strides=(2,2),kernel_initializer = 'he_normal')(b4)
    x= concatenate([conv11,conv8])
    conv12 = Conv2D(256,(3,3), activation = 'relu', padding = 'same', kernel_initializer = 'he_normal')(x)
    d6=Dropout(0.4)(conv12)
    conv13 = Conv2D(256,(3,3), activation = 'relu', padding = 'same', kernel_initializer = 'he_normal')(d6)
    b5=BatchNormalization()(conv13)
    
    
    conv14 = Conv2DTranspose(256,(4,4), activation = 'relu', padding = 'same', strides=(2,2),kernel_initializer = 'he_normal')(b5)
    x1=concatenate([conv14,conv6])
    conv15 = Conv2D(128,3, activation = 'relu', padding = 'same', kernel_initializer = 'he_normal')(x1)
    d7=Dropout(0.3)(conv15)
    conv16 = Conv2D(128,3, activation = 'relu', padding = 'same', kernel_initializer = 'he_normal')(d7)
    b6=BatchNormalization()(conv16)
    
    conv17 = Conv2DTranspose(128,(4,4), activation = 'relu', padding = 'same',strides=(2,2), kernel_initializer = 'he_normal')(b6)
    x2=concatenate([conv17,conv4])
    conv18 = Conv2D(64,(3,3), activation = 'relu', padding = 'same', kernel_initializer = 'he_normal')(x2)
    d8=Dropout(0.2)(conv18)
    conv19 = Conv2D(64,(3,3) ,activation = 'relu', padding = 'same', kernel_initializer = 'he_normal')(d8)
    b7=BatchNormalization()(conv19)
    
    conv20 = Conv2DTranspose(64,(4,4), activation = 'relu', padding = 'same',strides=(2,2), kernel_initializer = 'he_normal')(b7)
    x3=concatenate([conv20,conv2])
    conv21 = Conv2D(32,(3,3) ,activation = 'relu', padding = 'same', kernel_initializer = 'he_normal')(x3)
    d9=Dropout(0.1)(conv21)
    conv22 = Conv2D(32,(3,3), activation = 'relu', padding = 'same', kernel_initializer = 'he_normal')(d9)
    
    outputs = Conv2D(1,(1,1), activation = last_activation, padding = 'same', kernel_initializer = 'he_normal')(conv22)
    model2 = Model( inputs = inputs, outputs = outputs)
    
    return model2

#下载数据

from download_dataset import *
import os
path = "/content/Data"
if os.path.exists(path+'/DentalPanoramicXrays.zip') == False:
  os.mkdir(path)
  download_dataset(path+'/')


# 准备数据
from images_prepare import *
#pre_images(resize_shape,path,include_zip)
X,X_sizes=pre_images((512,512),path,True)


# 准备训练数据的遮罩
from masks_prepare import *
#Y=pre_masks(resize_shape=(512,512),path='/content/Segmentation-of-Teeth-in-Panoramic-X-ray-Image/Original_Masks')  ORIGINALL MASKS function 
Y=pre_splitted_masks(path='/content/Segmentation-of-Teeth-in-Panoramic-X-ray-Image/Custom_Masks') #Custom Splitted MASKS size 512x512

X=np.float32(X/255)
Y=np.float32(Y/255)

x_train=X[:105,:,:,:]
y_train=Y[:105,:,:,:]
x_test=X[105:,:,:,:]
y_test=Y[105:,:,:,:]

import random
import matplotlib.pyplot as plt
random_number=random.randint(0,104)
print(random_number)




# 模型预测
from model import *
model=UNET(input_shape=(512,512,1),last_activation='sigmoid')
model.summary()

model.compile(optimizer ='adam', loss = 'binary_crossentropy', metrics = ['accuracy'])
#Your choice batch and epoch 
model.fit(x_train,y_train,batch_size=8,epochs=150,verbose=1)

predict_img=model.predict(x_test)
##model.save(path)
predict=predict_img[1,:,:,0] 


plt.figure(figsize = (20,10))
plt.title("Predict Mask",fontsize = 40)
plt.imshow(predict)
#For CCA, we saved
plt.imsave("/content/predict.png",predict)





# 第二步修正结果
from google.colab.patches import cv2_imshow
import cv2
from CCA_Analysis import *


##Plotting - RESULT Example
img=cv2.imread("/content/Data/Images/107.png")#original img 107.png 

predict1 = cv2.resize(predict, (img.shape[1],img.shape[0]), interpolation=cv2.INTER_LANCZOS4)

mask=np.uint8(predict1*255)# 
_, mask = cv2.threshold(mask, thresh=255/2, maxval=255, type=cv2.THRESH_BINARY)
cnts,hieararch=cv2.findContours(mask,cv2.RETR_TREE,cv2.CHAIN_APPROX_SIMPLE)
img = cv2.drawContours(img[:,:,0], cnts, -1, (255, 0, 0) , 2)
img = cv2.UMat.get(img)
cv2_imshow(img)

from google.colab.patches import cv2_imshow
import cv2
from CCA_Analysis import *


##Plotting - RESULT Example
img=cv2.imread("/content/Data/Images/107.png")#original img 107.png 

predict1 = cv2.resize(predict, (img.shape[1],img.shape[0]), interpolation=cv2.INTER_LANCZOS4)

mask=np.uint8(predict1*255)# 
_, mask = cv2.threshold(mask, thresh=255/2, maxval=255, type=cv2.THRESH_BINARY)
cnts,hieararch=cv2.findContours(mask,cv2.RETR_TREE,cv2.CHAIN_APPROX_SIMPLE)
img = cv2.drawContours(img[:,:,0], cnts, -1, (255, 0, 0) , 2)
img = cv2.UMat.get(img)
cv2_imshow(img)