In [0]:

    

from matplotlib.pyplot import imshow
import numpy as np
from PIL import Image, ImageDraw, ImageFilter, ImageOps
from random import randint
import tensorflow as tf
import keras

from keras.models import Model, load_model
from keras.layers import Input, BatchNormalization, Activation, Dense, Dropout
from keras.layers.core import Lambda, RepeatVector, Reshape
from keras.layers.convolutional import Conv2D, Conv2DTranspose
from keras.layers.pooling import MaxPooling2D, GlobalMaxPool2D
from keras.layers.merge import concatenate, add
from keras.callbacks import EarlyStopping, ModelCheckpoint, ReduceLROnPlateau
from keras.optimizers import Adam
from keras.preprocessing.image import ImageDataGenerator, array_to_img, img_to_array, load_img


    

In [330]:

    

! git clone https://github.com/dreampulse/cell-count-experiment.git


    

    




fatal: destination path 'cell-count-experiment' already exists and is not an empty directory.

In [331]:

    

!ls cell-count-experiment/cell-images


    

    




Cells1BackgroundMasked.png  Cells1Mask.png  Cells2.jpg	    Cells3.jpg
Cells1Masked.png	    Cells1.png	    Cells2Mask.png  Cells3Mask.png

In [332]:

    

cells1 = Image.open("cell-count-experiment/cell-images/Cells1.png").convert('L')
cells2 = Image.open("cell-count-experiment/cell-images/Cells2.jpg").convert('L')
cells3 = Image.open("cell-count-experiment/cell-images/Cells3.jpg").convert('L')

basewidth = 1000
wpercent = (basewidth / float(cells1.size[0]))
hsize = int((float(cells1.size[1]) * float(wpercent)))
cells1 = cells1.resize((basewidth, hsize), Image.ANTIALIAS)
cells2 = cells2.resize((basewidth, hsize), Image.ANTIALIAS)
cells3 = cells3.resize((basewidth, hsize), Image.ANTIALIAS)

imshow(cells2)


    

    
Out[332]:





<matplotlib.image.AxesImage at 0x7f5ba29b2f28>






    

In [333]:

    

cells1_label = ImageOps.invert(Image.open("cell-count-experiment/cell-images/Cells1Mask.png").convert('L'))
cells2_label = ImageOps.invert(Image.open("cell-count-experiment/cell-images/Cells2Mask.png").convert('L'))
cells3_label = ImageOps.invert(Image.open("cell-count-experiment/cell-images/Cells3Mask.png").convert('L'))

cells1_label = cells1_label.resize((basewidth, hsize), Image.ANTIALIAS)
cells2_label = cells2_label.resize((basewidth, hsize), Image.ANTIALIAS)
cells3_label = cells3_label.resize((basewidth, hsize), Image.ANTIALIAS)

imshow(cells2_label)


    

    
Out[333]:





<matplotlib.image.AxesImage at 0x7f5ba29836d8>






    

In [334]:

    

tiles_size = 128

def get_cropped_img(img, i, j, tiles_size):
    return img.crop((i, j, i+tiles_size, j+tiles_size))

imshow(get_cropped_img(cells1, 100, 100, tiles_size))


    

    
Out[334]:





<matplotlib.image.AxesImage at 0x7f5ba2957a90>






    

In [335]:

    

def get_cropped_label(img, i, j, tiles_size):
  cropped_img = img.crop((i, j, i+tiles_size, j+tiles_size))
  # cropped_label = np.array(cropped_img)
  # center_pos = int(tiles_size/2)
  # prob_cell_in_center = cropped_label[center_pos, center_pos] / 255.0
  # return cropped_img, prob_cell_in_center, cropped_label
  return cropped_img

example_cropped_img = get_cropped_label(cells1_label, 100, 100, tiles_size)
imshow(example_cropped_img)


    

    
Out[335]:





<matplotlib.image.AxesImage at 0x7f5ba28b2ac8>






    

In [381]:

    

def create_data(inputs):
  x_training = []
  y_training = []

  def append_data(img, label_img, step_size = 40):
    for j in range(0, img.height - tiles_size, step_size):
      for i in range(0, img.width - tiles_size, step_size):
        x_training.append(np.array(get_cropped_img(img, i, j, tiles_size)) / 255.0)
        y_training.append(np.array(get_cropped_label(label_img, i, j, tiles_size)) / 255.0)

  for img, label_img in inputs:
     append_data(img, label_img)

  # return np.array(x_training), keras.utils.to_categorical(np.where(np.array(np.array(y_training)) > .5, 1, 0), num_classes=2)
  return np.array(x_training), np.array(y_training)

x_train, y_train = create_data([(cells1, cells1_label), (cells3, cells3_label)])
x_valid, y_valid = create_data([(cells2, cells2_label)])
print("Number of test images", len(x_train))
print("Number of validation images", len(x_valid))
# x_training[0]


    

    




Number of test images 704
Number of validation images 352

In [0]:

    

x_train = np.expand_dims(x_train, axis=3)
x_valid = np.expand_dims(x_valid, axis=3)

y_train = np.expand_dims(y_train, axis=3)
y_valid = np.expand_dims(y_valid, axis=3)


    

In [0]:

    

# x_train.reshape(x_train.shape[0], tiles_size, tiles_size,1)


    

In [0]:

    




    

In [366]:

    

x_train.shape


    

    
Out[366]:





(468, 128, 128, 1)

In [0]:

    

def conv2d_block(input_tensor, n_filters, kernel_size = 3, batchnorm = True):
  """Function to add 2 convolutional layers with the parameters passed to it"""
  # first layer
  x = Conv2D(filters = n_filters, kernel_size = (kernel_size, kernel_size),\
            kernel_initializer = 'he_normal', padding = 'same')(input_tensor)
  if batchnorm:
      x = BatchNormalization()(x)
  x = Activation('relu')(x)

  # second layer
  x = Conv2D(filters = n_filters, kernel_size = (kernel_size, kernel_size),\
            kernel_initializer = 'he_normal', padding = 'same')(input_tensor)
  if batchnorm:
      x = BatchNormalization()(x)
  x = Activation('relu')(x)

  return x

def get_unet(input_img, n_filters = 16, dropout = 0.1, batchnorm = True):
  """Function to define the UNET Model"""
  # Contracting Path
  c1 = conv2d_block(input_img, n_filters * 1, kernel_size = 3, batchnorm = batchnorm)
  p1 = MaxPooling2D((2, 2))(c1)
  p1 = Dropout(dropout)(p1)

  c2 = conv2d_block(p1, n_filters * 2, kernel_size = 3, batchnorm = batchnorm)
  p2 = MaxPooling2D((2, 2))(c2)
  p2 = Dropout(dropout)(p2)

  c3 = conv2d_block(p2, n_filters * 4, kernel_size = 3, batchnorm = batchnorm)
  p3 = MaxPooling2D((2, 2))(c3)
  p3 = Dropout(dropout)(p3)

  c4 = conv2d_block(p3, n_filters * 8, kernel_size = 3, batchnorm = batchnorm)
  p4 = MaxPooling2D((2, 2))(c4)
  p4 = Dropout(dropout)(p4)

  c5 = conv2d_block(p4, n_filters = n_filters * 16, kernel_size = 3, batchnorm = batchnorm)

  # Expansive Path
  u6 = Conv2DTranspose(n_filters * 8, (3, 3), strides = (2, 2), padding = 'same')(c5)
  u6 = concatenate([u6, c4])
  u6 = Dropout(dropout)(u6)
  c6 = conv2d_block(u6, n_filters * 8, kernel_size = 3, batchnorm = batchnorm)

  u7 = Conv2DTranspose(n_filters * 4, (3, 3), strides = (2, 2), padding = 'same')(c6)
  u7 = concatenate([u7, c3])
  u7 = Dropout(dropout)(u7)
  c7 = conv2d_block(u7, n_filters * 4, kernel_size = 3, batchnorm = batchnorm)

  u8 = Conv2DTranspose(n_filters * 2, (3, 3), strides = (2, 2), padding = 'same')(c7)
  u8 = concatenate([u8, c2])
  u8 = Dropout(dropout)(u8)
  c8 = conv2d_block(u8, n_filters * 2, kernel_size = 3, batchnorm = batchnorm)

  u9 = Conv2DTranspose(n_filters * 1, (3, 3), strides = (2, 2), padding = 'same')(c8)
  u9 = concatenate([u9, c1])
  u9 = Dropout(dropout)(u9)
  c9 = conv2d_block(u9, n_filters * 1, kernel_size = 3, batchnorm = batchnorm)

  outputs = Conv2D(1, (1, 1), activation='sigmoid')(c9)
  model = Model(inputs=[input_img], outputs=[outputs])
  return model


    

In [368]:

    

x_train[0].shape[0], x_train[0].shape[1], 1


    

    
Out[368]:





(128, 128, 1)

In [369]:

    

# model = keras.models.Sequential([
#     keras.layers.Conv2D(32, (5, 5), padding='same', activation='relu', input_shape=x_train.shape[1:]),
#     keras.layers.MaxPooling2D(pool_size=(3,3)), 
#     keras.layers.Conv2D(64, (5, 5), activation='relu'),
#     keras.layers.MaxPooling2D(pool_size=(4,4)), 
#     keras.layers.Dropout(0.5),    
#     keras.layers.Flatten(),
    
#     keras.layers.Dense(75*75, activation="sigmoid"),

#     keras.layers.Reshape((75,75))
# ])

input_img = keras.layers.Input((x_train[0].shape[0], x_train[0].shape[1], 1), name='img')
model = get_unet(input_img, n_filters=16, dropout=0.05, batchnorm=True)

model.summary()


    

    




Model: "model_9"
__________________________________________________________________________________________________
Layer (type)                    Output Shape         Param #     Connected to                     
==================================================================================================
img (InputLayer)                (None, 128, 128, 1)  0                                            
__________________________________________________________________________________________________
conv2d_207 (Conv2D)             (None, 128, 128, 16) 160         img[0][0]                        
__________________________________________________________________________________________________
batch_normalization_196 (BatchN (None, 128, 128, 16) 64          conv2d_207[0][0]                 
__________________________________________________________________________________________________
activation_196 (Activation)     (None, 128, 128, 16) 0           batch_normalization_196[0][0]    
__________________________________________________________________________________________________
max_pooling2d_55 (MaxPooling2D) (None, 64, 64, 16)   0           activation_196[0][0]             
__________________________________________________________________________________________________
dropout_86 (Dropout)            (None, 64, 64, 16)   0           max_pooling2d_55[0][0]           
__________________________________________________________________________________________________
conv2d_209 (Conv2D)             (None, 64, 64, 32)   4640        dropout_86[0][0]                 
__________________________________________________________________________________________________
batch_normalization_198 (BatchN (None, 64, 64, 32)   128         conv2d_209[0][0]                 
__________________________________________________________________________________________________
activation_198 (Activation)     (None, 64, 64, 32)   0           batch_normalization_198[0][0]    
__________________________________________________________________________________________________
max_pooling2d_56 (MaxPooling2D) (None, 32, 32, 32)   0           activation_198[0][0]             
__________________________________________________________________________________________________
dropout_87 (Dropout)            (None, 32, 32, 32)   0           max_pooling2d_56[0][0]           
__________________________________________________________________________________________________
conv2d_211 (Conv2D)             (None, 32, 32, 64)   18496       dropout_87[0][0]                 
__________________________________________________________________________________________________
batch_normalization_200 (BatchN (None, 32, 32, 64)   256         conv2d_211[0][0]                 
__________________________________________________________________________________________________
activation_200 (Activation)     (None, 32, 32, 64)   0           batch_normalization_200[0][0]    
__________________________________________________________________________________________________
max_pooling2d_57 (MaxPooling2D) (None, 16, 16, 64)   0           activation_200[0][0]             
__________________________________________________________________________________________________
dropout_88 (Dropout)            (None, 16, 16, 64)   0           max_pooling2d_57[0][0]           
__________________________________________________________________________________________________
conv2d_213 (Conv2D)             (None, 16, 16, 128)  73856       dropout_88[0][0]                 
__________________________________________________________________________________________________
batch_normalization_202 (BatchN (None, 16, 16, 128)  512         conv2d_213[0][0]                 
__________________________________________________________________________________________________
activation_202 (Activation)     (None, 16, 16, 128)  0           batch_normalization_202[0][0]    
__________________________________________________________________________________________________
max_pooling2d_58 (MaxPooling2D) (None, 8, 8, 128)    0           activation_202[0][0]             
__________________________________________________________________________________________________
dropout_89 (Dropout)            (None, 8, 8, 128)    0           max_pooling2d_58[0][0]           
__________________________________________________________________________________________________
conv2d_215 (Conv2D)             (None, 8, 8, 256)    295168      dropout_89[0][0]                 
__________________________________________________________________________________________________
batch_normalization_204 (BatchN (None, 8, 8, 256)    1024        conv2d_215[0][0]                 
__________________________________________________________________________________________________
activation_204 (Activation)     (None, 8, 8, 256)    0           batch_normalization_204[0][0]    
__________________________________________________________________________________________________
conv2d_transpose_38 (Conv2DTran (None, 16, 16, 128)  295040      activation_204[0][0]             
__________________________________________________________________________________________________
concatenate_38 (Concatenate)    (None, 16, 16, 256)  0           conv2d_transpose_38[0][0]        
                                                                 activation_202[0][0]             
__________________________________________________________________________________________________
dropout_90 (Dropout)            (None, 16, 16, 256)  0           concatenate_38[0][0]             
__________________________________________________________________________________________________
conv2d_217 (Conv2D)             (None, 16, 16, 128)  295040      dropout_90[0][0]                 
__________________________________________________________________________________________________
batch_normalization_206 (BatchN (None, 16, 16, 128)  512         conv2d_217[0][0]                 
__________________________________________________________________________________________________
activation_206 (Activation)     (None, 16, 16, 128)  0           batch_normalization_206[0][0]    
__________________________________________________________________________________________________
conv2d_transpose_39 (Conv2DTran (None, 32, 32, 64)   73792       activation_206[0][0]             
__________________________________________________________________________________________________
concatenate_39 (Concatenate)    (None, 32, 32, 128)  0           conv2d_transpose_39[0][0]        
                                                                 activation_200[0][0]             
__________________________________________________________________________________________________
dropout_91 (Dropout)            (None, 32, 32, 128)  0           concatenate_39[0][0]             
__________________________________________________________________________________________________
conv2d_219 (Conv2D)             (None, 32, 32, 64)   73792       dropout_91[0][0]                 
__________________________________________________________________________________________________
batch_normalization_208 (BatchN (None, 32, 32, 64)   256         conv2d_219[0][0]                 
__________________________________________________________________________________________________
activation_208 (Activation)     (None, 32, 32, 64)   0           batch_normalization_208[0][0]    
__________________________________________________________________________________________________
conv2d_transpose_40 (Conv2DTran (None, 64, 64, 32)   18464       activation_208[0][0]             
__________________________________________________________________________________________________
concatenate_40 (Concatenate)    (None, 64, 64, 64)   0           conv2d_transpose_40[0][0]        
                                                                 activation_198[0][0]             
__________________________________________________________________________________________________
dropout_92 (Dropout)            (None, 64, 64, 64)   0           concatenate_40[0][0]             
__________________________________________________________________________________________________
conv2d_221 (Conv2D)             (None, 64, 64, 32)   18464       dropout_92[0][0]                 
__________________________________________________________________________________________________
batch_normalization_210 (BatchN (None, 64, 64, 32)   128         conv2d_221[0][0]                 
__________________________________________________________________________________________________
activation_210 (Activation)     (None, 64, 64, 32)   0           batch_normalization_210[0][0]    
__________________________________________________________________________________________________
conv2d_transpose_41 (Conv2DTran (None, 128, 128, 16) 4624        activation_210[0][0]             
__________________________________________________________________________________________________
concatenate_41 (Concatenate)    (None, 128, 128, 32) 0           conv2d_transpose_41[0][0]        
                                                                 activation_196[0][0]             
__________________________________________________________________________________________________
dropout_93 (Dropout)            (None, 128, 128, 32) 0           concatenate_41[0][0]             
__________________________________________________________________________________________________
conv2d_223 (Conv2D)             (None, 128, 128, 16) 4624        dropout_93[0][0]                 
__________________________________________________________________________________________________
batch_normalization_212 (BatchN (None, 128, 128, 16) 64          conv2d_223[0][0]                 
__________________________________________________________________________________________________
activation_212 (Activation)     (None, 128, 128, 16) 0           batch_normalization_212[0][0]    
__________________________________________________________________________________________________
conv2d_224 (Conv2D)             (None, 128, 128, 1)  17          activation_212[0][0]             
==================================================================================================
Total params: 1,179,121
Trainable params: 1,177,649
Non-trainable params: 1,472
__________________________________________________________________________________________________

In [0]:

    

model.compile(loss="binary_crossentropy",
#               optimizer='sgd',
              # optimizer='adam',
              optimizer=keras.optimizers.Adadelta(),
              # metrics=["binary_accuracy"]
              metrics=["accuracy"]
              )


    

In [423]:

    

history = model.fit(x_train, y_train, epochs=5, validation_data=(x_valid, y_valid))


    

    




Train on 704 samples, validate on 352 samples
Epoch 1/5
704/704 [==============================] - 4s 6ms/step - loss: 0.1421 - binary_accuracy: 0.6039 - val_loss: 0.1734 - val_binary_accuracy: 0.7423
Epoch 2/5
704/704 [==============================] - 4s 6ms/step - loss: 0.1420 - binary_accuracy: 0.6039 - val_loss: 0.1771 - val_binary_accuracy: 0.7421
Epoch 3/5
704/704 [==============================] - 4s 6ms/step - loss: 0.1421 - binary_accuracy: 0.6040 - val_loss: 0.1800 - val_binary_accuracy: 0.7400
Epoch 4/5
704/704 [==============================] - 4s 6ms/step - loss: 0.1424 - binary_accuracy: 0.6039 - val_loss: 0.1772 - val_binary_accuracy: 0.7406
Epoch 5/5
704/704 [==============================] - 4s 6ms/step - loss: 0.1422 - binary_accuracy: 0.6039 - val_loss: 0.1772 - val_binary_accuracy: 0.7418

In [424]:

    

import pandas as pd
import matplotlib.pyplot as plt

pd.DataFrame(history.history).plot(figsize=(8, 5))
plt.grid(True)
plt.gca().set_ylim(0, 1) # set the vertical range to [0-1]
plt.show()


    

In [473]:

    

# Das hier ist ein Trainingsbild
img_index = 2

x_org = x_valid.reshape(x_valid.shape[0], tiles_size, tiles_size)[img_index]*255
imshow(Image.fromarray(x_org.astype(np.uint8)))


    

    
Out[473]:





<matplotlib.image.AxesImage at 0x7f5b4a584400>






    

In [0]:

    




    

In [474]:

    

# Das hier sind die Labels die trainiert wurden
y_org =  y_valid.reshape(y_valid.shape[0], tiles_size, tiles_size)[img_index]*255
imshow(Image.fromarray(y_org.astype(np.uint8)))


    

    
Out[474]:





<matplotlib.image.AxesImage at 0x7f5b4a5dd6d8>






    

In [475]:

    

# Das hier hat das netzt predicted
y_pred = model.predict(np.array([x_valid[img_index]]))[0].reshape(tiles_size, tiles_size)*255
imshow(Image.fromarray(y_pred.astype(np.uint8)))


    

    
Out[475]:





<matplotlib.image.AxesImage at 0x7f5b49b7de10>






    

In [0]:

    




    

In [0]:

    

width_tiles = int(cells2.width/tiles_size)
height_tiles = int(cells2.height/tiles_size)


    

In [0]:

    

# make it dividable by tiles
val_img = cells2.crop((0,0,width_tiles*tiles_size, height_tiles*tiles_size))
val = np.array(val_img) / 255.0


    

In [0]:

    

imgs = []

for j in range(height_tiles):
  for i in range(width_tiles):
    imgs.append(val[j*tiles_size:(j*tiles_size)+tiles_size, i*tiles_size:(i*tiles_size)+tiles_size])

imgs = np.array(imgs).reshape(-1, tiles_size, tiles_size, 1)

# Image.fromarray((imgs[0] * 255).astype(np.uint8))

pred_imgs = model.predict(imgs)
# pred_imgs = imgs


    

In [353]:

    

img = np.zeros([height_tiles*tiles_size, width_tiles*tiles_size])

n = 0
for j in range(height_tiles):
  for i in range(width_tiles):
    img[j*tiles_size:(j*tiles_size)+tiles_size, i*tiles_size:(i*tiles_size)+tiles_size] = pred_imgs[n].reshape(tiles_size, tiles_size)
    n = n + 1


Image.fromarray((img * 255).astype(np.uint8))


    

    
Out[353]:

In [354]:

    

val_img


    

    
Out[354]:

In [0]: