3DUNetCNN Pretrained Models

Pre-trained model and configuration files designed to work with the 3DUNetCNN GitHub project.

The model files are named: <dataset>_<receptive_field>[_<version>][_<fold_number>].h5

The corresponding configuration files are also given. These files include important details regarding data preprocessing, model architecture, and model training. The easiest way to train one of these models to on a new application is to modify the given configuration file to suit your needs and then to feed configuration file along with the pre-trained model to the "train.py" script from the GitHub repository. The receptive field is given for reference for those looking to implement a model. However, taking a pre-trained model at one receptive field and then training at another receptive field is perfectly reasonable, and will like reduce training time compared to training model from scratch. It is also worth noting that the project employs flexible model loading that allows for the weights from pretrained UNet models to be used as weights for models with slightly different UNet architectures. Therefore, even if you change architecture (using the configuration file), you should still be able to load weights from a pretrained model with a different architecture to reduce the training time needed.

When using the AutoImplant models, please cite the following paper:

• Ellis D.G., Aizenberg M.R. (2020) Deep Learning Using Augmentation via Registration: 1st Place Solution to the AutoImplant 2020 Challenge. In: Li J., Egger J. (eds) Towards the Automatization of Cranial Implant Design in Cranioplasty. AutoImplant 2020. Lecture Notes in Computer Science, vol 12439. Springer, Cham. https://doi.org/10.1007/978-3-030-64327-0_6

The augmented AutoImplant data used to train the models can be found here.