FlyByCNN - pytorch3d implementation

This is a new implementation of fly-by-cnn using the pytorch3D framework.
This implementation also provides the possibility to output the dental labels using the Universal Labeling System or the World Dental Federation notation (FDI).
The model used by this release was trained using the data from the teeth segmentation challenge from MICCAI.
More information about the challenge and implementation of the 3DSlicer extension here

ULMS - Universal Labeling and Merging

The release of this application was done at https://github.com/RomainUSA/fly-by-cnn the tag is 2.0.3
The models for the UMLMS labeling are at https://github.com/RomainUSA/fly-by-cnn/releases/tag/2.0.3

FlyByCNN - Dilate boundary as parameter

The dilate is optional now as a parameter for hard cases that do not separate the boundary between GUM and Teeth well.
This model is also trained with better ground truth and has the best accuracy so far.

FlyByCNN - Dilate boundary during prediction

This release includes a new feature to dilate the boundary between the teeth and gum to split them more accurately.
The UNET used here is trained using upper and lower arches rotated randomly.
The training uses weights - [0.5, 1, 1, 4] (background, gum, teeth, boundary).

FlyByCNN - upper and lower arches

This release includes a trained model for the segmentation of lower and upper arches.
The model is trained using 40 lower arches models and 38 upper arches.
Each model is randomly rotated 2 times. An icosahedron sampling with subdivision level 2 is used (42 points of view). A total of 9829 image sample pairs are extracted.

FlyBy with point Id maps

This release modifies the mechanism to sample the 3D objects by implementing point id maps.
It also includes a docker file that can be used to create a container that will have all the requirements to run fly by CNN.

FlyBy CNN and Dental model seg

This is the first release for the dental model seg application.

Tools versions:
python3.6
Tensorflow2.2

The trained model is Segmented_meshes_train_rotated_samples_nc4.tar.gz, it is saved in the saved_model format using tensorflow1.15.