An intelligent postoperative management system for glioblastoma integrating automated segmentation, risk stratification, and recurrence spatial mapping

Purpose

This study aimed to develop DeepGBM-Recure, an integrated artificial intelligence (AI) system for optimizing precision radiotherapy and individualized surveillance in glioblastoma (GBM) by automating postoperative risk stratification and spatial targeting of recurrence hotspots.

Methods

This DeepGBM-Recure system comprises three synergistic modules: 1) Automated segmentation of peri-cavitary hyperintense regions on postoperative fluid-attenuated inversion recovery (FLAIR) images using a 3D nnU-Net framework; 2) Patient-level early recurrence prediction based on radiomics features and random forest classification; 3) Voxel-wise spatial mapping of high-risk subregions via supervoxel analysis. The system was trained and validated on data from 145 patients across two centers and externally tested on data from 39 patients across another two centers.

Results

On the test set, the nnU-Net segmentation model achieved a mean Dice coefficient of 0.85 ± 0.09. The patient-level and voxel-level prediction models achieved area under the ROC curves (AUCs) of 0.76 and 0.80, respectively. Notably, the voxel-level model exhibited strong spatial concordance between predicted high-risk heatmaps and ground-truth recurrence regions. Performance was further supported by calibration curves, decision curve analysis, and clinical application in representative cases, demonstrating favorable predictive accuracy in real-world scenarios.

Conclusion

DeepGBM-Recure represents a pioneering integrated solution that combines automated anatomical delineation, individualized risk stratification, and spatial recurrence guidance, offering a clinically applicable tool for precision radiotherapy and individualized surveillance. Prospective multi-center trials with larger cohorts are warranted to validate clinical utility and facilitate integration into real-world decision-making workflows.