ID #508 Scalable 3D longitudinal tumor monitoring in pediatric low-grade glioma using a Slicer-embedded nnU-Net
Nour Eltaani, Tahsin Ahmed, Aloys Portafaix, Samuel Kadoury, Dorsa Sadat, Mathieu Dehaes, Sébastien PerreaultAbstract
Background
Pediatric low-grade glioma (PLGG) responses are typically measured using 2D assessments, but irregular lesion morphology presents a significant clinical challenge. 3D volumetric analysis could improve monitoring, but manual segmentation is labor-intensive and limits scalability.
Methods
We analyzed MRIs from a clinical trial (TRAM-01-NCT03363217) of patients with refractory or recurrent PLGG treated with trametinib for 18 months. We developed a 3D Slicer extension utilizing a 3D full-resolution nnU-Net architecture. The model was trained on 132 expert-confirmed T2-FLAIR studies of PLGG with BRAF fusion. To evaluate clinical utility, we tested a personalized longitudinal strategy: for a patient’s timeline, n scans were used for training and m reserved for testing. This approach allows the model to learn patient-specific anatomical baselines to ensure consistent longitudinal tracking.
Results
A total of 445 MRIs were annotated. For a representative patient, training on a subset of the available scans and testing on the rest yielded high predictive value (Dice score 0.95–0.98), indicating strong intra-patient utility. Preliminary testing on an unseen PLGG subgroup (n = 17) not included in training showed strong generalization (mean Dice ≈ 0.78; IQR 0.66–0.88). Automation significantly improved workflow, reducing median processing time from around 25 minutes per scan manually to 1–2 minutes with the tool.
Conclusion
A Slicer-integrated nnU-Net provides a practical, expert-supervised solution for scalable 3D longitudinal volumetry. This demonstrates that full-resolution models effectively capture patient-specific progression. Future work will extend training to all TRAM-01 groups and implement a longitudinal strategy using prior segmentations as a reference to improve predictions for a follow-up study and inference time.