This repository will be used to test different approach for disc labeling
NeuroPoly disc labeling implementations:
- Hourglass approach: https://github.com/spinalcordtoolbox/disc-labeling-hourglass
- nnU-Net approach: https://github.com/spinalcordtoolbox/disc-labeling-nnunet
- Disc labeling benchmark: https://github.com/spinalcordtoolbox/disc-labeling-benchmark
This work is based on this paper and this implementation from MONAI.
The objective of this project is to reconstruct full MRI scans of the spine from partial FOV using latent diffusion models.
The training was done using ~60 stiched T2w MRI scans of healthy patients. The models used were a 2d LDM and a 2d VQ-VAE model (used to work in the latent space). To extend the number of images, 5 centered slices were extracted from each T2w scans.
In this first result, a cervical scan from a dataset unseen during training was gradually added during the diffusion process (inference) to condition the network to reconstruct the rest of the body. Each of the five images represent a different slice in the 3d input scan (the middle image corresponds to the middle slice of the 3d scan).
This other image shows another result when a lumbar scan is used as an input.
Too few images were available for the training of the LDM.