The repository contains 3D variants of popular models for segmentation like FPN, Unet, Linknet and PSPNet.

This repository is based on great segmentation_models repo by @qubvel

• Unet
• FPN
• Linknet
• PSPNet

• tensorflow>=2.8.0
• keras_applications>=1.0.8
• classification_models_3D>=1.0.6

pip install segmentation-models-3D

import segmentation_models_3D as sm

model1 = sm.Unet(
    'resnet34', 
    encoder_weights='imagenet'
)

# binary segmentation (this parameters are default when you call Unet('resnet34')
model2 = sm.FPN(
    'densenet121', 
    classes=1, 
    activation='sigmoid'
)

# multiclass segmentation with non overlapping class masks (your classes + background)
model3 = sm.Linknet(
    'resnet34', 
    classes=3, 
    activation='softmax'
)

# multiclass segmentation with independent overlapping/non-overlapping class masks
model4 = sm.PSPNet(
    'resnet34', 
    classes=3,
    activation='sigmoid'
)

# If you need to specify non-standard input shape
model5 = sm.Unet(
    'resnet50', 
    input_shape=(96, 128, 128, 6), 
    encoder_weights=None
)

All possible backbones: 'resnet18, 'resnet34', 'resnet50', 'resnet101', 'resnet152', 'seresnet18', 'seresnet34', 'seresnet50', 'seresnet101', 'seresnet152', 'seresnext50', 'seresnext101', 'senet154', 'resnext50', 'resnext101', 'vgg16', 'vgg19', 'densenet121', 'densenet169', 'densenet201', 'inceptionresnetv2', 'inceptionv3', 'mobilenet', 'mobilenetv2', 'efficientnetb0', 'efficientnetb1', 'efficientnetb2', 'efficientnetb3', 'efficientnetb4', 'efficientnetb5', 'efficientnetb6', 'efficientnetb7', 'efficientnetv2-b1', 'efficientnetv2-b2', 'efficientnetv2-b3', 'efficientnetv2-s', 'efficientnetv2-m', 'efficientnetv2-l'

More examples can be found in tst_keras.py

There is training example in training_example.py

• I tried to keep code as simple as possible
• I couldn't find good dataset for 3D segmentation task. So I randomly generate 3D volumes with dark background with light figures (spheres and cuboids) and model tries to segment these figures independetly. 1st mask for circles and 2nd mask for cuboids.

• Add stride_size parameter for better control of models

• https://github.com/qubvel/classification_models - original classification 2D repo
• https://github.com/qubvel/segmentation_models - original segmentation 2D repo
• classification_models_3D - models for classification in 3D
• volumentations - 3D augmentations

• There is no 'bilinear' interpolation for UpSample3D layer, so it uses Nearest Neighbour upsampling.

For more details, please refer to the publication: https://doi.org/10.1016/j.compbiomed.2021.105089

If you find this code useful, please cite it as:

@article{solovyev20223d,
  title={3D convolutional neural networks for stalled brain capillary detection},
  author={Solovyev, Roman and Kalinin, Alexandr A and Gabruseva, Tatiana},
  journal={Computers in Biology and Medicine},
  volume={141},
  pages={105089},
  year={2022},
  publisher={Elsevier},
  doi={10.1016/j.compbiomed.2021.105089}
}