Abstract Background In the postoperative setting, distinguishing between treatment-related changes and glioma recurrence remains a major challenge in neuroradiology. This pilot study explored a dedicated radiomics-based approach applied to postoperative MRI. Methods We retrospectively analyzed postoperative MRI scans from 38 patients with diffuse adult-type gliomas treated at the University Hospital Zurich. Recurrence was confirmed histologically or by clinico-radiological consensus. New contrast-enhancing VOIs were manually delineated on T1c images, and corresponding radiomics features were extracted from T1c and FLAIR images and were aggregated across VOIs. Logistic regression models were trained and evaluated on repeated random splits, with feature importance assessed through univariable analyses and selection frequency across iterations. Results Feature aggregation by minimum value (min) and selection by maximum volume (maxVol) produced significantly predictive models for both FLAIR- and T1c-derived features. Shape features achieved significantly predictive performance, with AUCs of 0.78 0.70,0.85 (FLAIR) and 0.76 0.70,0.81 (T1c) on internal testing. Among individual features, Shape Sphericity showed the highest discriminative ability, distinguishing recurrence with AUCs of 0.79 (maxVol aggregation) and 0.82 (min aggregation). Conclusion We identified robust shape-based radiomics features on T1c imaging that were significantly predictive of glioma recurrence. Dominant features—including Sphericity, Volume, and Surface-to-Volume Ratio—were consistently selected in 70%–100% of modeling repetitions. Pending prospective validation in larger multi-institutional cohorts, these findings may ultimately support more accurate clinical decision-making in postoperative glioma management.
Abler et al. (Tue,) studied this question.