This repository contains 3D variants of popular classification CNN models like ResNets, DenseNets, VGG, etc for keras module. It also contains weights obtained by converting ImageNet weights from the same 2D models.

This repository is based on great classification_models repo by @qubvel

• VGG [16, 19]
• ResNet [18, 34, 50, 101, 152]
• ResNeXt [50, 101]
• SE-ResNet [18, 34, 50, 101, 152]
• SE-ResNeXt [50, 101]
• SE-Net [154]
• DenseNet [121, 169, 201]
• Inception ResNet V2
• Inception V3
• MobileNet
• MobileNet v2
• EfficientNet [B0, B1, B2, B3, B4, B5, B6, B7]
• EfficientNet v2 [B0, B1, B2, B3, S, M, L]
• ConvNeXt

pip install classification-models-3D

from classification_models_3D.kkeras import Classifiers

ResNet18, preprocess_input = Classifiers.get('resnet18')
model = ResNet18(input_shape=(128, 128, 128, 3), weights='imagenet')

Keras 3 support different backends like: Tensorflow, Torch and Jax. Below you can find examples for different backends:

• Tensorflow: tst_keras_tensorflow.py
• Tensorflow: tst_special_cases_keras_tensorflow.py
• Torch: tst_keras_torch.py
• Torch: tst_special_cases_keras_torch.py
• Jax: tst_keras_jax.py
• Jax: tst_special_cases_keras_jax.py

• Tensorflow: training_example_keras_tensorflow.py
• Torch: training_example_keras_torch.py
• Jax: training_example_keras_jax.py

'resnet18', 'resnet34', 'resnet50', 'resnet101', 'resnet152', 'seresnet18', 'seresnet34', 'seresnet50', 'seresnet101', 'seresnet152', 'seresnext50', 'seresnext101', 'senet154', 'resnext50', 'resnext101', 'vgg16', 'vgg19', 'densenet121', 'densenet169', 'densenet201', 'mobilenet', 'mobilenetv2', 'inceptionresnetv2', 'inceptionv3', 'efficientnetb0', 'efficientnetb1', 'efficientnetb2', 'efficientnetb3', 'efficientnetb4', 'efficientnetb5', 'efficientnetb6', 'efficientnetb7', 'efficientnetv2-b0', 'efficientnetv2-b1', 'efficientnetv2-b2', 'efficientnetv2-b3', 'efficientnetv2-s', 'efficientnetv2-m', 'efficientnetv2-l', 'convnext_tiny', 'convnext_small', 'convnext_base', 'convnext_large', 'convnext_xlarge'

Code to convert 2D imagenet weights to 3D variant is available here: convert_imagenet_weights_to_3D_models.py.

If initial 2D model had shape (512, 512, 3) then you can use shape (D, H, W, 3) where D * H * W ~= 512*512, so something like (64, 64, 64, 3) will be ok.

Training with single NVIDIA 1080Ti (11 GB) worked with:

• DenseNet121, DenseNet169 and ResNet50 with shape (96, 128, 128, 3) and batch size 6
• DenseNet201 with shape (96, 128, 128, 3) and batch size 5
• ResNet18 with shape (128, 160, 160, 3) and batch size 6

Default pooling/stride size for 3D models is set equal to 2. You can change it for your needs using parameter stride_size. Example:

from classification_models_3D.kkeras import Classifiers

ResNet18, preprocess_input = Classifiers.get('resnet18')
model = ResNet18(
   input_shape=(224, 224, 224, 3),
   stride_size=4,
   kernel_size=3, 
   weights=None
)

stride_size can be:

• single integer. Example: 4
• tuple of size 5 (if you didn't change repetition parameter). Example: (2, 2, 4, 2, 2)
• tuple of tuples. Example: ( (2, 2, 1), (2, 2, 4), (2, 2, 2), (2, 1, 2), (2, 4, 2), ). Each number in (2, 2, 1) control stride of individual dimension.

• For some models like (resnet, resnext, senet, vgg16, vgg19, densenet) it's possible to change number of blocks/poolings. For example if you want to make more poolings overall. You can do it like that:

from classification_models_3D.kkeras import Classifiers

ResNet18, preprocess_input = Classifiers.get('resnet18')
model = ResNet18(
   input_shape=(128, 128, 128, 3),
   include_top=False,
   weights=None,
   stride_size=(1, 1, 2, 2, 2, 2, 2, 2),
   repetitions=(2, 2, 2, 2, 2, 2, 2),
   init_filters=16,
)

• Note 1: Since number of filters grows 2 times, you can set initial number of filters with init_filters parameter.
• Note 2: There is no imagenet weights for models which were modified this way.

• https://github.com/qubvel/classification_models - original 2D repo
• timm_3d - models for classification in 3D for PyTorch
• segmentation models 3D - models for segmentation in 3D for Keras/Tensorflow
• volumentations - 3D augmentations
• driven_data_repo - code for training and inference on real dataset

Last version which supports Keras2 is 1.0.10

pip install classification-models-3D==1.0.10

• There is no DepthwiseConv3D layer in keras, so repo used custom layer from this repo by @alexandrosstergiou which can be slower than native implementation.
• There is no imagenet weights for 'inceptionresnetv2' and 'inceptionv3'.

This code was used to get 1st place in DrivenData: Advance Alzheimer’s Research with Stall Catchers competition.

More details on ArXiv: https://arxiv.org/abs/2104.01687

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}
}