A deep learning–based model for postoperative resection assessment in glioblastoma: a comparative validation study

Background Post-operative imaging of glioblastoma presents unique challenges for tumor segmentation due to surgical cavities, hemorrhage, and treatment-related changes. Although multiple open-source artificial intelligence (AI) tools have demonstrated strong performance in pre-operative settings, their utility in post-operative assessment has been insufficiently validated. Method Our newly developed subtraction-based brain tumor segmentation model (Dynapex BT) was tested on the LUMIERE dataset, a publicly available cohor. Its performance was evaluated in three domains: (1) extent of resection (EOR) classification by experts, (2) segmentation performance for residual tumors with non-gross total resection, and (3) correlation between EOR classification and measured residual tumor volume. Performance of our model was compared with that of DeepBraTumIA, a widely used commercially available U-Net-based brain tumor segmentation tool. Results Sixty-eight post-operative MRI scans from the LUMIERE dataset met the inclusion criteria. Dynapex BT demonstrated higher accuracy in EOR classification compared to DeepBraTumIA (sensitivity 0.60, specificity 1.00 vs. sensitivity 0.23, specificity 0.93 for gross total resection). The area under curve was significantly higher for Dynapex BT (0.826 vs. 0.659, p  = 0.024). Dynapex BT also achieved significantly better segmentation performance compared to DeepBraTumIA (DSC: 0.815 vs. 0.406, p  = 0.002; precision: 0.771 vs. 0.366, p  < 0.001; recall: 0.888 vs. 0.583, p  = 0.019). Conclusion Our automated postoperative segmentation model outperformed a widely used commercial U-Net-based model not only in segmentation accuracy but also in clinically relevant endpoints. Future studies in larger, multi-institutional cohorts are warranted to evaluate its clinical utility.