time-distance-transformer

Implementation of vision transformers for sparse, irregularly sampled imaging. Please cite the preprint if you find this helpful: Time-distance vision transformers in lung cancer diagnosis from longitudinal computed tomography

Prequisite datasets and resources

• NLST dataset: This model was validated using an imaging cohort from the National Lung Screening Trial. Lung screening CTs are available through the NIH: https://cdas.cancer.gov/nlst/
• tumor_cifar: Synthetic dataset used as proof of concept: https://github.com/MASILab/tumor-cifar
• Pulmonary nodule detection from Liao et al.[1]: Model that proposes pulmonary nodule ROIs. Both the raw ROIs or feature vectors from this algorithm are used as input to our model: https://github.com/lfz/DSB2017

Training

1. Install dependencies from requirements.txt.
2. Edit config files in config/*.YAML to point to data location.
   • Generate image paths and store as dictionary with each entry being a cross validation fold.
   • see paper for choices of positional and time embeddings
3. Run main_nlst.py
   • If input is whole image or ROI, masked autoencoder pretraining is necessary: python main_nlst.py --config nlst.YAML --pretrain --kfold 0
   • If input is feature vectors, pretrainin is NOT needed: python main_nlst.py --config nlst.YAML --train --kfold 0

References

1. Liao F, Liang M, Li Z, Hu X, Song S. Evaluate the Malignancy of Pulmonary Nodules Using the 3D Deep Leaky Noisy-or Network. IEEE Transactions on Neural Networks and Learning Systems. 2017;30(11):3484-3495. doi:10.1109/tnnls.2019.2892409.

---

The contents covered by this repository, including code and pretrained models in the docker container, are free for noncommercial usage (CC BY-NC 4.0). Please check the LICENSE.md file for more details of the copyright information.