| Sumario: | Simple Summary: After chemoradiotherapy for HNSCC, cervical lymph nodes persist despite treatment, yet only a subset harbor active tumors. Identifying vital from non-vital nodes on CT scans helps guide decisions. We developed a machine learning pipeline that extracts image features from CT scans and tested whether augmenting training data with realistic transformations improves classification. This strategy improved classification performance by up to 21.9%, although differences did not reach statistical significance in this limited sample. Background: Distinguishing vital from non-vital persistent cervical lymph nodes after chemoradiotherapy in HNSCC remains clinically challenging. We investigated whether image-level data augmentation improves CT-based radiomics classification for this task. Methods: We evaluated eight augmentation strategies and their 28 pairwise combinations in 55 patients, using Bayesian hyperparameter tuning with Optuna for parameter optimization. A radiomics pipeline comprising Radiomics features, five feature selectors, and seven classifiers was assessed using patient-level stratified 5-fold cross-validation. Configurations were ranked using a composite score defined as the mean of AUC, ACC and F1-score. Results: Feature selection improved the composite score from 0.659 to 0.742. The best augmented configuration, Window Contrast Variation, achieved a composite score of 0.803 and an AUC of 0.831, corresponding to an 8.2% relative point-estimate gain over feature selection alone and a 21.9% gain over the no-selection baseline when feature selection and augmentation were combined. Conclusions: These findings suggest that feature selection with optimized augmentation may enhance radiomics-based lymph node classification. However, individual augmentation-versus-baseline differences did not reach statistical significance in this limited sample, requiring confirmation in larger cohorts.
|
|---|