This repository performs survival analysis using clinical and radiomics features.
It utilizes two commonly used packages for computing radiomic features and survival model, pyradiomics and pysurvival.
https://pyradiomics.readthedocs.io/en/latest/
https://square.github.io/pysurvival/
Survival analysis predicts when an event is likely to happen. It is essentially a regression model but with censored events. Censoring refers to the cases that the event has not occurred. A regression model would be biased toward the cases that did experience the event. Therefore, the input is features (X), time to event (T), and event (E). Event is a boolean that indicates whether it occurred or not.
Performance of a survival model is usually characterized by C-index and/or Integrated Brier Score.
These are assumed to be categorical and converted to one-hot.
Radiomics converts unstructured data (images) into structured data (case x features). They are pre-computed using pyradiomics in batch mode due to speed.
The stability/reproducibilty is not well studied or poor due to the varying hardware and scanning parameters. We established stability by Test-Retest experiments with imaging within two weeks. Unstable radiomic features outside given p-value range is discarded.
Additionally, features can be filtered with other feature selection methods. We used LASSO with bootstrap and determined a cutoff for predictiveness.
The final output is a csv file with each row as a model and the columns as its model performance.
02/25/2020 Uploaded