Author: Hao Peng, Yajun Yu 👨🔬
Affiliation: Department of Radiation Oncology, UT Southwestern Medical Center, Medical Artificial Intelligence and Automation (MAIA) Lab, Department of Radiation Oncology, UT Southwestern Medical Center 🌍
Purpose: Personalized ultra-fractionated stereotactic adaptive radiotherapy (PULSAR) is a new treatment paradigm pioneered by our institution. But the early decision-making process in PULSAR is challenging by means of a data-driven approach, due to small and imbalanced dataset. While feature selection in radiomics is well-established, our study presents a novel approach that distinguishes itself from traditional methods, such as least absolute shrinkage and selection operator (Lasso) technique, enabling more accurate prediction models.
Methods: Leveraging the principles of compressed sensing (CS), we developed two CS-based feature selection techniques (binary and Gaussian random projections) and applied them to a brain metastasis case, consisting of 69 lesions, treated with PULSAR. Unlike Lasso, which relies on a single projection matrix, CS models benefit from ensembling through various binary or Gaussian random projection patterns.
Results: Our findings show that CS-based approaches outperform the widely used Lasso. A combination of residual error and frequency-based feature selection outperforms weight coefficient-based selection criteria. For instance, with the 5-feature sets, the performance metrics between CS-Binary model and Lasso model with “RE-cnt” criterion are as follows: AUC (0.937 vs. 0.890), balanced accuracy (87.3% vs. 83.0%), and F1 score (79.4% vs. 75.9%).
Conclusion: Our proposed framework is able to simplify feature selection, enhance predictive accuracy, and support early decision-making in personalized radiotherapy. Our methods present unique advantages in two practical scenarios. First, clinical trials often begin with small datasets, which complicates radiomics analysis. Second, it is crucial to select the most informative features while minimizing their number to enhance interpretability.