Impact of sampling strategies on the classification of micro-FTIR hyperspectral data

Abstract

This study evaluated class balancing strategies for classifying oral squamous cell carcinoma (OSCC) in FTIR hyperspectral images using the XGBoost model. Although the dataset was balanced at the image level, spectral quality filtering introduced pixel-level class imbalance. Resampling methods— SMOTE, Tomek Links, and their combination—were tested, as well as AllKNN for redundancy reduction. All approaches outperformed the unbalanced baseline, but the best overall performance metrics were achieved with the combined use of AllKNN and Tomek Links.