This model was used in my bachelor thesis for segmenting bacteria in blood. The dataset and an example Jupyter Notebook can be found at the Kaggle Bacteria detection with darkfield microscopy.
Following are some results of the thesis.
Manual segmentation as ground truth. Blood cells tagged with red color, bacteria tagged with yellow color.
 |
 |
 |
 |
 |
 |
 |
 |
 |
 |
 |
 |
UNet++ is a new general purpose image segmentation architecture for more accurate image segmentation. UNet++ consists of U-Nets of varying depths whose decoders are densely connected at the same resolution via the redesigned skip pathways, which aim to address two key challenges of the U-Net: 1) unknown depth of the optimal architecture and 2) the unnecessarily restrictive design of skip connections.
Following is a visual comparision of Unet (left) and Unet++ (right).
 |
 |
Clone the repository and import the model with:
from unetpp import model as unetpp
from unet import model as unet
from tensorflow.keras.optimizers import Adam
from losses import dice, iou, weighted_loss, cce_iou_dice
weights_list = {1: 1.0, 2: 50.0, 3: 70.0}
if __name__ == '__main__':
optimizer = Adam(lr=0.0005)
unet.compile(optimizer=optimizer, loss=weighted_loss(cce_iou_dice, weights_list), metrics=[dice, iou])
unet.summary()
unetpp.compile(optimizer=optimizer, loss=weighted_loss(cce_iou_dice, weights_list), metrics=[dice, iou])
unetpp.summary()This model was designed for multi-classes classification and therefore uses softmax
as final activation. In order to fight class imbalance, which is usually the case, the high-order function
model.losses.weighted_loss can be used to decorate classical loss functions.
- Sorence Dice
- Crossentropy
- F1
Kera's summary of model with 32 convolution filters at the input.
__________________________________________________________________________________________________
Layer (type) Output Shape Param # Connected to
==================================================================================================
input_1 (InputLayer) [(None, 512, 512, 51 0
__________________________________________________________________________________________________
conv2d (Conv2D) (None, 512, 512, 32) 147488 input_1[0][0]
__________________________________________________________________________________________________
batch_normalization (BatchNorma (None, 512, 512, 32) 128 conv2d[0][0]
__________________________________________________________________________________________________
leaky_re_lu (LeakyReLU) (None, 512, 512, 32) 0 batch_normalization[0][0]
__________________________________________________________________________________________________
dropout (Dropout) (None, 512, 512, 32) 0 leaky_re_lu[0][0]
__________________________________________________________________________________________________
conv2d_1 (Conv2D) (None, 512, 512, 32) 9248 dropout[0][0]
__________________________________________________________________________________________________
batch_normalization_1 (BatchNor (None, 512, 512, 32) 128 conv2d_1[0][0]
__________________________________________________________________________________________________
leaky_re_lu_1 (LeakyReLU) (None, 512, 512, 32) 0 batch_normalization_1[0][0]
__________________________________________________________________________________________________
dropout_1 (Dropout) (None, 512, 512, 32) 0 leaky_re_lu_1[0][0]
__________________________________________________________________________________________________
max_pooling2d (MaxPooling2D) (None, 256, 256, 32) 0 dropout_1[0][0]
__________________________________________________________________________________________________
conv2d_2 (Conv2D) (None, 256, 256, 64) 18496 max_pooling2d[0][0]
__________________________________________________________________________________________________
batch_normalization_2 (BatchNor (None, 256, 256, 64) 256 conv2d_2[0][0]
__________________________________________________________________________________________________
leaky_re_lu_2 (LeakyReLU) (None, 256, 256, 64) 0 batch_normalization_2[0][0]
__________________________________________________________________________________________________
dropout_2 (Dropout) (None, 256, 256, 64) 0 leaky_re_lu_2[0][0]
__________________________________________________________________________________________________
conv2d_3 (Conv2D) (None, 256, 256, 64) 36928 dropout_2[0][0]
__________________________________________________________________________________________________
batch_normalization_3 (BatchNor (None, 256, 256, 64) 256 conv2d_3[0][0]
__________________________________________________________________________________________________
leaky_re_lu_3 (LeakyReLU) (None, 256, 256, 64) 0 batch_normalization_3[0][0]
__________________________________________________________________________________________________
dropout_3 (Dropout) (None, 256, 256, 64) 0 leaky_re_lu_3[0][0]
__________________________________________________________________________________________________
max_pooling2d_1 (MaxPooling2D) (None, 128, 128, 64) 0 dropout_3[0][0]
__________________________________________________________________________________________________
conv2d_6 (Conv2D) (None, 128, 128, 128 73856 max_pooling2d_1[0][0]
__________________________________________________________________________________________________
batch_normalization_6 (BatchNor (None, 128, 128, 128 512 conv2d_6[0][0]
__________________________________________________________________________________________________
leaky_re_lu_6 (LeakyReLU) (None, 128, 128, 128 0 batch_normalization_6[0][0]
__________________________________________________________________________________________________
dropout_5 (Dropout) (None, 128, 128, 128 0 leaky_re_lu_6[0][0]
__________________________________________________________________________________________________
conv2d_7 (Conv2D) (None, 128, 128, 128 147584 dropout_5[0][0]
__________________________________________________________________________________________________
batch_normalization_7 (BatchNor (None, 128, 128, 128 512 conv2d_7[0][0]
__________________________________________________________________________________________________
leaky_re_lu_7 (LeakyReLU) (None, 128, 128, 128 0 batch_normalization_7[0][0]
__________________________________________________________________________________________________
dropout_6 (Dropout) (None, 128, 128, 128 0 leaky_re_lu_7[0][0]
__________________________________________________________________________________________________
max_pooling2d_2 (MaxPooling2D) (None, 64, 64, 128) 0 dropout_6[0][0]
__________________________________________________________________________________________________
conv2d_12 (Conv2D) (None, 64, 64, 256) 295168 max_pooling2d_2[0][0]
__________________________________________________________________________________________________
batch_normalization_12 (BatchNo (None, 64, 64, 256) 1024 conv2d_12[0][0]
__________________________________________________________________________________________________
leaky_re_lu_12 (LeakyReLU) (None, 64, 64, 256) 0 batch_normalization_12[0][0]
__________________________________________________________________________________________________
dropout_9 (Dropout) (None, 64, 64, 256) 0 leaky_re_lu_12[0][0]
__________________________________________________________________________________________________
conv2d_13 (Conv2D) (None, 64, 64, 256) 590080 dropout_9[0][0]
__________________________________________________________________________________________________
batch_normalization_13 (BatchNo (None, 64, 64, 256) 1024 conv2d_13[0][0]
__________________________________________________________________________________________________
leaky_re_lu_13 (LeakyReLU) (None, 64, 64, 256) 0 batch_normalization_13[0][0]
__________________________________________________________________________________________________
dropout_10 (Dropout) (None, 64, 64, 256) 0 leaky_re_lu_13[0][0]
__________________________________________________________________________________________________
max_pooling2d_3 (MaxPooling2D) (None, 32, 32, 256) 0 dropout_10[0][0]
__________________________________________________________________________________________________
conv2d_20 (Conv2D) (None, 32, 32, 512) 1180160 max_pooling2d_3[0][0]
__________________________________________________________________________________________________
batch_normalization_20 (BatchNo (None, 32, 32, 512) 2048 conv2d_20[0][0]
__________________________________________________________________________________________________
leaky_re_lu_20 (LeakyReLU) (None, 32, 32, 512) 0 batch_normalization_20[0][0]
__________________________________________________________________________________________________
conv2d_21 (Conv2D) (None, 32, 32, 512) 2359808 leaky_re_lu_20[0][0]
__________________________________________________________________________________________________
batch_normalization_21 (BatchNo (None, 32, 32, 512) 2048 conv2d_21[0][0]
__________________________________________________________________________________________________
leaky_re_lu_21 (LeakyReLU) (None, 32, 32, 512) 0 batch_normalization_21[0][0]
__________________________________________________________________________________________________
dropout_14 (Dropout) (None, 32, 32, 512) 0 leaky_re_lu_21[0][0]
__________________________________________________________________________________________________
conv2d_transpose_6 (Conv2DTrans (None, 64, 64, 256) 524544 dropout_14[0][0]
__________________________________________________________________________________________________
concatenate_6 (Concatenate) (None, 64, 64, 512) 0 conv2d_transpose_6[0][0]
dropout_10[0][0]
__________________________________________________________________________________________________
conv2d_transpose_3 (Conv2DTrans (None, 128, 128, 32) 32800 dropout_10[0][0]
__________________________________________________________________________________________________
conv2d_22 (Conv2D) (None, 64, 64, 256) 1179904 concatenate_6[0][0]
__________________________________________________________________________________________________
concatenate_3 (Concatenate) (None, 128, 128, 160 0 dropout_6[0][0]
conv2d_transpose_3[0][0]
__________________________________________________________________________________________________
conv2d_transpose_1 (Conv2DTrans (None, 256, 256, 32) 16416 dropout_6[0][0]
__________________________________________________________________________________________________
batch_normalization_22 (BatchNo (None, 64, 64, 256) 1024 conv2d_22[0][0]
__________________________________________________________________________________________________
conv2d_14 (Conv2D) (None, 128, 128, 32) 46112 concatenate_3[0][0]
__________________________________________________________________________________________________
concatenate_1 (Concatenate) (None, 256, 256, 96) 0 dropout_3[0][0]
conv2d_transpose_1[0][0]
__________________________________________________________________________________________________
conv2d_transpose (Conv2DTranspo (None, 512, 512, 32) 8224 dropout_3[0][0]
__________________________________________________________________________________________________
leaky_re_lu_22 (LeakyReLU) (None, 64, 64, 256) 0 batch_normalization_22[0][0]
__________________________________________________________________________________________________
batch_normalization_14 (BatchNo (None, 128, 128, 32) 128 conv2d_14[0][0]
__________________________________________________________________________________________________
conv2d_8 (Conv2D) (None, 256, 256, 32) 27680 concatenate_1[0][0]
__________________________________________________________________________________________________
concatenate (Concatenate) (None, 512, 512, 64) 0 dropout_1[0][0]
conv2d_transpose[0][0]
__________________________________________________________________________________________________
conv2d_23 (Conv2D) (None, 64, 64, 256) 590080 leaky_re_lu_22[0][0]
__________________________________________________________________________________________________
leaky_re_lu_14 (LeakyReLU) (None, 128, 128, 32) 0 batch_normalization_14[0][0]
__________________________________________________________________________________________________
batch_normalization_8 (BatchNor (None, 256, 256, 32) 128 conv2d_8[0][0]
__________________________________________________________________________________________________
conv2d_4 (Conv2D) (None, 512, 512, 32) 18464 concatenate[0][0]
__________________________________________________________________________________________________
batch_normalization_23 (BatchNo (None, 64, 64, 256) 1024 conv2d_23[0][0]
__________________________________________________________________________________________________
conv2d_15 (Conv2D) (None, 128, 128, 32) 9248 leaky_re_lu_14[0][0]
__________________________________________________________________________________________________
leaky_re_lu_8 (LeakyReLU) (None, 256, 256, 32) 0 batch_normalization_8[0][0]
__________________________________________________________________________________________________
batch_normalization_4 (BatchNor (None, 512, 512, 32) 128 conv2d_4[0][0]
__________________________________________________________________________________________________
leaky_re_lu_23 (LeakyReLU) (None, 64, 64, 256) 0 batch_normalization_23[0][0]
__________________________________________________________________________________________________
batch_normalization_15 (BatchNo (None, 128, 128, 32) 128 conv2d_15[0][0]
__________________________________________________________________________________________________
conv2d_9 (Conv2D) (None, 256, 256, 32) 9248 leaky_re_lu_8[0][0]
__________________________________________________________________________________________________
leaky_re_lu_4 (LeakyReLU) (None, 512, 512, 32) 0 batch_normalization_4[0][0]
__________________________________________________________________________________________________
dropout_15 (Dropout) (None, 64, 64, 256) 0 leaky_re_lu_23[0][0]
__________________________________________________________________________________________________
leaky_re_lu_15 (LeakyReLU) (None, 128, 128, 32) 0 batch_normalization_15[0][0]
__________________________________________________________________________________________________
batch_normalization_9 (BatchNor (None, 256, 256, 32) 128 conv2d_9[0][0]
__________________________________________________________________________________________________
conv2d_5 (Conv2D) (None, 512, 512, 32) 9248 leaky_re_lu_4[0][0]
__________________________________________________________________________________________________
conv2d_transpose_7 (Conv2DTrans (None, 128, 128, 128 131200 dropout_15[0][0]
__________________________________________________________________________________________________
dropout_11 (Dropout) (None, 128, 128, 32) 0 leaky_re_lu_15[0][0]
__________________________________________________________________________________________________
leaky_re_lu_9 (LeakyReLU) (None, 256, 256, 32) 0 batch_normalization_9[0][0]
__________________________________________________________________________________________________
batch_normalization_5 (BatchNor (None, 512, 512, 32) 128 conv2d_5[0][0]
__________________________________________________________________________________________________
concatenate_7 (Concatenate) (None, 128, 128, 288 0 conv2d_transpose_7[0][0]
dropout_6[0][0]
dropout_11[0][0]
__________________________________________________________________________________________________
dropout_7 (Dropout) (None, 256, 256, 32) 0 leaky_re_lu_9[0][0]
__________________________________________________________________________________________________
conv2d_transpose_4 (Conv2DTrans (None, 256, 256, 32) 4128 dropout_11[0][0]
__________________________________________________________________________________________________
leaky_re_lu_5 (LeakyReLU) (None, 512, 512, 32) 0 batch_normalization_5[0][0]
__________________________________________________________________________________________________
conv2d_24 (Conv2D) (None, 128, 128, 128 331904 concatenate_7[0][0]
__________________________________________________________________________________________________
concatenate_4 (Concatenate) (None, 256, 256, 128 0 dropout_3[0][0]
dropout_7[0][0]
conv2d_transpose_4[0][0]
__________________________________________________________________________________________________
dropout_4 (Dropout) (None, 512, 512, 32) 0 leaky_re_lu_5[0][0]
__________________________________________________________________________________________________
conv2d_transpose_2 (Conv2DTrans (None, 512, 512, 32) 4128 dropout_7[0][0]
__________________________________________________________________________________________________
batch_normalization_24 (BatchNo (None, 128, 128, 128 512 conv2d_24[0][0]
__________________________________________________________________________________________________
conv2d_16 (Conv2D) (None, 256, 256, 32) 36896 concatenate_4[0][0]
__________________________________________________________________________________________________
concatenate_2 (Concatenate) (None, 512, 512, 96) 0 dropout_1[0][0]
dropout_4[0][0]
conv2d_transpose_2[0][0]
__________________________________________________________________________________________________
leaky_re_lu_24 (LeakyReLU) (None, 128, 128, 128 0 batch_normalization_24[0][0]
__________________________________________________________________________________________________
batch_normalization_16 (BatchNo (None, 256, 256, 32) 128 conv2d_16[0][0]
__________________________________________________________________________________________________
conv2d_10 (Conv2D) (None, 512, 512, 32) 27680 concatenate_2[0][0]
__________________________________________________________________________________________________
conv2d_25 (Conv2D) (None, 128, 128, 128 147584 leaky_re_lu_24[0][0]
__________________________________________________________________________________________________
leaky_re_lu_16 (LeakyReLU) (None, 256, 256, 32) 0 batch_normalization_16[0][0]
__________________________________________________________________________________________________
batch_normalization_10 (BatchNo (None, 512, 512, 32) 128 conv2d_10[0][0]
__________________________________________________________________________________________________
batch_normalization_25 (BatchNo (None, 128, 128, 128 512 conv2d_25[0][0]
__________________________________________________________________________________________________
conv2d_17 (Conv2D) (None, 256, 256, 32) 9248 leaky_re_lu_16[0][0]
__________________________________________________________________________________________________
leaky_re_lu_10 (LeakyReLU) (None, 512, 512, 32) 0 batch_normalization_10[0][0]
__________________________________________________________________________________________________
leaky_re_lu_25 (LeakyReLU) (None, 128, 128, 128 0 batch_normalization_25[0][0]
__________________________________________________________________________________________________
batch_normalization_17 (BatchNo (None, 256, 256, 32) 128 conv2d_17[0][0]
__________________________________________________________________________________________________
conv2d_11 (Conv2D) (None, 512, 512, 32) 9248 leaky_re_lu_10[0][0]
__________________________________________________________________________________________________
dropout_16 (Dropout) (None, 128, 128, 128 0 leaky_re_lu_25[0][0]
__________________________________________________________________________________________________
leaky_re_lu_17 (LeakyReLU) (None, 256, 256, 32) 0 batch_normalization_17[0][0]
__________________________________________________________________________________________________
batch_normalization_11 (BatchNo (None, 512, 512, 32) 128 conv2d_11[0][0]
__________________________________________________________________________________________________
conv2d_transpose_8 (Conv2DTrans (None, 256, 256, 64) 32832 dropout_16[0][0]
__________________________________________________________________________________________________
dropout_12 (Dropout) (None, 256, 256, 32) 0 leaky_re_lu_17[0][0]
__________________________________________________________________________________________________
leaky_re_lu_11 (LeakyReLU) (None, 512, 512, 32) 0 batch_normalization_11[0][0]
__________________________________________________________________________________________________
concatenate_8 (Concatenate) (None, 256, 256, 192 0 conv2d_transpose_8[0][0]
dropout_3[0][0]
dropout_7[0][0]
dropout_12[0][0]
__________________________________________________________________________________________________
dropout_8 (Dropout) (None, 512, 512, 32) 0 leaky_re_lu_11[0][0]
__________________________________________________________________________________________________
conv2d_transpose_5 (Conv2DTrans (None, 512, 512, 32) 4128 dropout_12[0][0]
__________________________________________________________________________________________________
conv2d_26 (Conv2D) (None, 256, 256, 64) 110656 concatenate_8[0][0]
__________________________________________________________________________________________________
concatenate_5 (Concatenate) (None, 512, 512, 128 0 dropout_1[0][0]
dropout_4[0][0]
dropout_8[0][0]
conv2d_transpose_5[0][0]
__________________________________________________________________________________________________
batch_normalization_26 (BatchNo (None, 256, 256, 64) 256 conv2d_26[0][0]
__________________________________________________________________________________________________
conv2d_18 (Conv2D) (None, 512, 512, 32) 36896 concatenate_5[0][0]
__________________________________________________________________________________________________
leaky_re_lu_26 (LeakyReLU) (None, 256, 256, 64) 0 batch_normalization_26[0][0]
__________________________________________________________________________________________________
batch_normalization_18 (BatchNo (None, 512, 512, 32) 128 conv2d_18[0][0]
__________________________________________________________________________________________________
conv2d_27 (Conv2D) (None, 256, 256, 64) 36928 leaky_re_lu_26[0][0]
__________________________________________________________________________________________________
leaky_re_lu_18 (LeakyReLU) (None, 512, 512, 32) 0 batch_normalization_18[0][0]
__________________________________________________________________________________________________
batch_normalization_27 (BatchNo (None, 256, 256, 64) 256 conv2d_27[0][0]
__________________________________________________________________________________________________
conv2d_19 (Conv2D) (None, 512, 512, 32) 9248 leaky_re_lu_18[0][0]
__________________________________________________________________________________________________
leaky_re_lu_27 (LeakyReLU) (None, 256, 256, 64) 0 batch_normalization_27[0][0]
__________________________________________________________________________________________________
batch_normalization_19 (BatchNo (None, 512, 512, 32) 128 conv2d_19[0][0]
__________________________________________________________________________________________________
dropout_17 (Dropout) (None, 256, 256, 64) 0 leaky_re_lu_27[0][0]
__________________________________________________________________________________________________
leaky_re_lu_19 (LeakyReLU) (None, 512, 512, 32) 0 batch_normalization_19[0][0]
__________________________________________________________________________________________________
conv2d_transpose_9 (Conv2DTrans (None, 512, 512, 32) 8224 dropout_17[0][0]
__________________________________________________________________________________________________
dropout_13 (Dropout) (None, 512, 512, 32) 0 leaky_re_lu_19[0][0]
__________________________________________________________________________________________________
concatenate_9 (Concatenate) (None, 512, 512, 160 0 conv2d_transpose_9[0][0]
dropout_1[0][0]
dropout_4[0][0]
dropout_8[0][0]
dropout_13[0][0]
__________________________________________________________________________________________________
conv2d_28 (Conv2D) (None, 512, 512, 32) 46112 concatenate_9[0][0]
__________________________________________________________________________________________________
batch_normalization_28 (BatchNo (None, 512, 512, 32) 128 conv2d_28[0][0]
__________________________________________________________________________________________________
leaky_re_lu_28 (LeakyReLU) (None, 512, 512, 32) 0 batch_normalization_28[0][0]
__________________________________________________________________________________________________
conv2d_29 (Conv2D) (None, 512, 512, 32) 9248 leaky_re_lu_28[0][0]
__________________________________________________________________________________________________
batch_normalization_29 (BatchNo (None, 512, 512, 32) 128 conv2d_29[0][0]
__________________________________________________________________________________________________
leaky_re_lu_29 (LeakyReLU) (None, 512, 512, 32) 0 batch_normalization_29[0][0]
__________________________________________________________________________________________________
dropout_18 (Dropout) (None, 512, 512, 32) 0 leaky_re_lu_29[0][0]
__________________________________________________________________________________________________
conv2d_30 (Conv2D) (None, 512, 512, 3) 99 dropout_18[0][0]
==================================================================================================
Total params: 8,340,483
Trainable params: 8,333,827
Non-trainable params: 6,656