3702 - Integrating Compressed Sensing with Radiomics Feature Selection for Outcome Prediction in Pulsar

Purpose/Objective(s): Personalized ultra-fractionated stereotactic adaptive radiotherapy (PULSAR) is an innovative treatment paradigm. However, early decision-making in PULSAR is challenging due to small and imbalanced dataset, posing challenges in feature selection and outcome prediction. Our study introduces a novel Compressed Sensing (CS)-based approach that distinguishes itself from conventional radiomics methods, such as least absolute shrinkage and selection operator (Lasso) technique.

Materials/Methods: Leveraging the CS principles, we developed two feature selection models (binary and Gaussian random projections) and applied them to a brain metastasis case, consisting of 69 lesions treated with a non-invasive stereotactic radiosurgery instrument and PULSAR. Once the top features were identified, support vector machine (SVM) models were tested to classify whether lesions exhibited =20% volume reduction at 3-month follow up. To address the limitations of small and imbalanced dataset, a stratified 5-fold cross-validation procedure was conducted with 10 repeats.

Results: Two CS-based models outperform the widely used Lasso. Unlike Lasso, which relies on a single projection matrix, CS models utilize ensemble learning through various binary or Gaussian random projection patterns, enhancing feature selection robustness. Combining residual error as the selection criteria and frequency-based feature selection shows improved performance over traditional weight coefficient-based criteria. With the 5-feature sets, the performance comparison 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%), F1 score (79.4% vs. 75.9%), Kappa coefficient (75.8% vs. 69.0%), and Matthews correlation coefficient (78.0% vs. 72.1%). “RE-cnt” criterion tends to select features with moderate correlations, but with strongest correlations with respect to the relative gross tumor volume (GTV) changes.

Conclusion: The CS-based framework helps streamline feature selection, enhance predictive accuracy, and early decision-making in PULSAR. It would be particularly advantageous in two practical scenarios: 1) clinical trials with small-sized datasets when a standard radiomics analysis is prone to overfitting, and 2) selecting features of highest prognostics values to enhance interpretability. By integrating CS-based feature selection and SVM classification, we can shift the decision-making process in adaptive radiotherapy from empirical judgments to a data-driven approach, enabling more personalized and effective treatment strategies.