MedNext-Insight Model for Automated Metabolic Tumor Volume Delineation on Computed Tomography and Prognostic Value in Nasopharyngeal Carcinoma - PubMed
6 hours ago
- #deep learning
- #nasopharyngeal carcinoma
- #automated delineation
- The MedNext-Insight deep learning model was developed to automatically delineate metabolic tumor volume (MTV) on routine computed tomography (CT) scans without needing positron emission tomography (PET), specifically for nasopharyngeal carcinoma (NPC).
- Using CT alone, the model was trained on a retrospective cohort of 392 NPC patients and outperformed other models (nnUNetV2, Pix2Pix, and 3D-CycleGAN) in segmentation accuracy with a Dice similarity coefficient (DSC) of 0.808 ± 0.110.
- Radiomic features extracted from predicted MTV showed strong reproducibility and prognostic performance comparable to ground truth MTV for event-free survival (EFS), with a concordance index (C-index) of 0.712 vs. 0.744.
- MTV-derived radiomics outperformed primary gross tumor volume (GTVp)-based features, especially when combined with clinical variables, achieving a C-index of 0.809.
- Internal temporal validation confirmed stable segmentation accuracy and prognostic discrimination using CT-only predicted MTV, demonstrating robustness across different scanners.
- The model offers a resource-efficient approach for risk stratification in NPC and supports potential future biology-guided adaptive radiotherapy, eliminating the need for PET imaging.