Quantum-Inspired Potential Mapping for Multi-Parametric MRI Radiomics–Based Post-Resection Glioblastoma Survival Prediction
Abstract
Purpose
To develop a multi-parametric MRI (mp-MRI) radiomics framework for predicting post-resection glioblastoma (GBM) survival by integrating conventional MR modalities with a quantum mechanics–inspired imaging representation.
Methods
A public dataset (n=236) with pre-operative T1, T1ce, T2, and FLAIR images was studied. A potential field (V), modeling the distribution of voxel-wise multi-modality intensity vectors within the tumor ROI, was computed for every patient. V was designed to capture common versus rare mp-MRI patterns and provide an additional image contrast. Radiomic features (first-order intensity, texture, and shape) were extracted from mp-MRI and V. Non-imaging clinical variables (age, resection method, and volumes of tumor subregions) were incorporated via positional encoding to improve fusion with radiomic features. Feature selection was performed using LASSO, with hyperparameters tuned via nested cross-validation. Selected features were used to train a support vector machine (SVM) classifier for survival prediction (short-term vs long-term). Two feature sets were compared: (a) Clinical + mpMRI and (b) Clinical + mpMRI + V.
Results
Across outer validation folds, adding V improved AUC from 0.656±0.097 to 0.673±0.087. On the held-out test set, the potential-augmented model achieved higher ROCAUC (0.72 vs 0.68) and accuracy (0.71 vs 0.63), with consistent improvements in precision, recall, specificity, and F1 score.
Conclusion
Quantum-inspired potential fields provide complementary prognostic information beyond conventional radiomics and clinical variables for post-resection GBM survival prediction.