我正在使用U-net模型进行MRImage分割。当我尝试运行它时，我希望能得到预测的图像，但是模型的精度却显示为0.99549。当我从预测数组绘制预测图像时，它什么也没显示。
K.set_image_data_format('channels_last') # TF dimension
_, *input_image_shape, _ = Xtr.shape
input_image_shape = tuple(input_image_shape)
smooth = 1.
use_dropout = True #FOR OVERFITTING
use_regularizers = True # overfitting
dropout_rate = 0.5
number_of_epochs = 40
batch_size = 32
kernel_size = (5, 5)
initial_volume_size = 64
def dice_coef_per_image_in_batch(y_true, y_pred):
y_true_f = K.batch_flatten(y_true)
y_pred_f = K.batch_flatten(y_pred)
intersection = 2. * K.sum(y_true_f * y_pred_f, axis=1, keepdims=True) + smooth
union = K.sum(y_true_f, axis=1, keepdims=True) + K.sum( y_pred_f, axis=1, keepdims=True) + smooth
return K.mean(intersection / union)
def dice_coef_loss(y_true, y_pred):
return -dice_coef_per_image_in_batch(y_true, y_pred)
def dice_coef_accur(y_true, y_pred):
return dice_coef_per_image_in_batch(y_true, y_pred)
def setup_regularizers(conv_layer):
return BatchNormalization()(conv_layer) if use_regularizers else conv_layer
def setup_dropout(conv_layer):
return Dropout(dropout_rate)(conv_layer) if use_dropout else conv_layer
inputs = Input((*input_image_shape, 1))
conv1 = Conv2D(initial_volume_size, kernel_size, activation='relu', padding='same')(inputs)
conv1 = Conv2D(initial_volume_size, kernel_size, activation='relu', padding='same')(conv1)
conv1 = setup_regularizers(conv1)
pool1 = MaxPooling2D(pool_size=(2, 2))(conv1)
conv2 = Conv2D(initial_volume_size*2, kernel_size, activation='relu', padding='same')(pool1)
conv2 = Conv2D(initial_volume_size*2, kernel_size, activation='relu', padding='same')(conv2)
conv2 = setup_regularizers(conv2)
pool2 = MaxPooling2D(pool_size=(2, 2))(conv2)
conv3 = Conv2D(initial_volume_size*4,kernel_size, activation='relu', padding='same')(pool2)
conv3 = Conv2D(initial_volume_size*4, kernel_size, activation='relu', padding='same')(conv3)
conv3 = setup_regularizers(conv3)
pool3 = MaxPooling2D(pool_size=(2, 2))(conv3)
conv4 = Conv2D(initial_volume_size*8, kernel_size, activation='relu', padding='same')(pool3)
conv4 = Conv2D(initial_volume_size*8, kernel_size, activation='relu', padding='same')(conv4)
conv4 = setup_regularizers(conv4)
pool4 = MaxPooling2D(pool_size=(2, 2))(conv4)
conv5 = Conv2D(initial_volume_size*16, kernel_size, activation='relu', padding='same')(pool4)
conv5 = Conv2D(initial_volume_size*16, kernel_size, activation='relu', padding='same')(conv5)
conv5 = setup_regularizers(conv5)
up6 = concatenate([UpSampling2D(size=(2, 2))(conv5), conv4], axis=3)
up6 = setup_dropout(up6)
conv6 = Conv2D(initial_volume_size*8, kernel_size, activation='relu', padding='same')(up6)
conv6 = Conv2D(initial_volume_size*8, kernel_size, activation='relu', padding='same')(conv6)
up7 = concatenate([UpSampling2D(size=(2, 2))(conv6), conv3], axis=3)
up7 = setup_dropout(up7)
conv7 = Conv2D(initial_volume_size*4, kernel_size, activation='relu', padding='same')(up7)
conv7 = Conv2D(initial_volume_size*4,kernel_size, activation='relu', padding='same')(conv7)
up8 = concatenate([UpSampling2D(size=(2, 2))(conv7), conv2], axis=3)
up8 = setup_dropout(up8)
conv8 = Conv2D(initial_volume_size*2, kernel_size, activation='relu', padding='same')(up8)
conv8 = Conv2D(initial_volume_size*2, kernel_size, activation='relu', padding='same')(conv8)
up9 = concatenate([UpSampling2D(size=(2, 2))(conv8), conv1], axis=3)
up9 = setup_dropout(up9)
conv9 = Conv2D(initial_volume_size, kernel_size, activation='relu', padding='same')(up9)
conv9 = Conv2D(initial_volume_size, kernel_size, activation='relu', padding='same')(conv9)
conv10 = Conv2D(1, (1, 1), activation='sigmoid')(conv9)
model = Model(inputs=[inputs], outputs=[conv10])
model.compile(optimizer=Adam(lr=1e-5), loss=dice_coef_loss, metrics=[dice_coef_accur])
print("Size of the CNN: %s" % model.count_params())
print(model.summary())
model_checkpoint = ModelCheckpoint( saved_model_filename, monitor='val_dice_coef_accur',save_best_only=True, verbose=1 )
csv_logger = CSVLogger(csv_logger_training, append=True, separator=';')
history = model.fit(Xtr, ytr, batch_size=batch_size, epochs=number_of_epochs, verbose=2, shuffle=True,
callbacks=[model_checkpoint, csv_logger], validation_data=(Xva, yva))
test_loss, accuracy_test = model.evaluate(Xte, yte, verbose=0)
print("Test Accuracy Mean: "+str(accuracy_test))
model.load_weights("datasets/dataset1/test-3.hdf5")
print(model.metrics_names)
test_loss, accuracy_test = model.evaluate(Xte, yte, verbose=0)
print("Test Accuracy Mean: "+str(accuracy_test))
imgs_mask_test = model.predict(Xte, verbose=1)
ncols = 3 # number of columns in final grid of images
nrows = 30 # looking at all images takes some time
_, axes = plt.subplots(nrows, ncols, figsize=(17, 17*nrows/ncols))
for axis in axes.flatten():
axis.set_axis_off()
axis.set_aspect('equal')
for k in range(0, nrows):
im_test_original = Xte[k].reshape(*input_image_shape)
im_result = imgs_mask_test[k].reshape(*input_image_shape)
im_ground_truth = yte[k].reshape(*input_image_shape)
axes[k, 0].set_title("Original Test Image")
axes[k, 0].imshow(im_test_original, cmap='gray')
axes[k, 1].set_title("Ground Truth")
axes[k, 1].imshow(im_ground_truth, cmap='gray')
axes[k, 2].set_title("Predicted")
axes[k, 2].imshow(im_result, cmap='gray')
如果有人可以告诉我我做错了什么，那将非常有帮助。