Image registration is one of the most critical problems in radiology targeting to establish correspondences between images modalities of the same patient or longitudinal studies. This problem is traditionally casted as an optimization problem. In the advent of deep learning, the objective of this project will be to study recent advances for unsupervised deep learning deformable registration in the context of CT images for radiation oncology.
Based on the LUNA16 dataset, we propose to use deep registration architectures to evaluate how leveraging segmentation data can help improve performances.
Prediction evolution through epochs, see report for more details
├── bin
│ ├── experiment_1
│ ├── experiment_2
│ └── experiment_3
├── data
├── docs
│ ├── img
│ ├── reports
│ └── readings
├── notebooks
│ ├── sandbox
│ └── clean
│ └── stats_extract.py
├── src
│ ├── evaluation
│ ├── generators
│ ├── layers
│ ├── networks
│ │ └── networks_utils
│ └── training
└── utils
bin: experiments directorydata: contains MRI-image datasets and segmentation masksdocs: useful papers, report and othernotebooks: useful scripts and notebooks of ongoing worksrc: contains all modules useful to run an experimentutils: miscellaneous utilities
