Abstract

Hypoxia plays a pivotal role in tumor aggressiveness and therapy resistance. We propose an interpretable machine learning pipeline for classifying hypoxia using oxygen maps simulated over synthetic 3D vascular networks. Oxygen distributions were modeled via reaction-diffusion equations with MichaelisMenten kinetics, solved using finite difference methods to produce PO2 maps across 159 vascular configurations. Maps were divided into 16 x 16 x 16 voxel patches, and 120 texture features were extracted per patch. Our methodology employed group-aware cross-validation, systematic class imbalance evaluation, and rigorous feature selection using univariate analysis and RFECV. Logistic Regression achieved near-perfect classification (F1-score: 0.9887 on held-out test set) with the complete feature set identified as near-optimal. Remarkably, performance plateaued at 45 features (F1-score: 0.9893), indicating that a compact feature subset captures most of the discriminative information. Interpretability analysis using SHAP and LIME consistently identified waveletbased features as the most influential predictors, with SHAP values of 7.83, 6.84, and 6.83 for sagittal, axial, and coronal wavelet coefficients, respectively. This pipeline enables accurate and interpretable hypoxia classification, establishing a foundation for non-invasive tumor oxygenation analysis.