International journal of radiation oncology, biology, physics · 2026
MedNext-Insight Model for Automated Metabolic Tumor Volume Delineation on Computed Tomography and Prognostic Value in Nasopharyngeal Carcinoma.
СтатьяНазвание и аннотация: Перевод пока не подготовлен
Обработка статьи · этапы, ошибки и управление
Загрузка состояния…
Ошибки и ограничения
История · было → стало
Состояние обновляется каждые 5 секунд, пока блок открыт. Процент показывается только там, где известен объём работы. Пауза останавливает продолжение этапа; текущий запрос к источнику может завершиться.
Hao Meng-Yu, Xiong Yu-Xi, Li Yu, Mo Zi-Jie, Feng Chen-Yang, Hu Jiang, Zhang Su-Man, Yang Yu-Xian, Jia Le-Cheng, Li Hua, He Yi-Da, Sun Xiao-Qing, Xu Sha, Zhou Guan-Qun, Sun Ying
Аннотация
PURPOSES: To develop a deep learning model for automated metabolic tumor volume (MTV) delineation on routine computed tomography (CT) without positron emission tomography (PET) and to validate its prognostic value in nasopharyngeal carcinoma (NPC).
METHODS AND MATERIALS: A retrospective cohort of 392 patients with NPC undergoing pre-radiotherapy 2-deoxy-2-[fluorine-18]fluoro-D-glucose PET/CT in 2021 was enrolled and randomly divided into training (n = 314, including 63 for validation) and test (n = 78) cohorts. Ground-truth MTV (GT_MTV) was generated from PET-registered CT using standardized uptake value SUV > 2.5 within the primary gross tumor volume. A 7-layer MedNext-Insight model with dual-window CT inputs and Dice-Focal loss was trained to predict MTV using CT alone. Segmentation performance was compared with no-new-U-Net version 2 (nnUNetV2), Conditional Generative Adversarial Network for Image-to-Image Translation (Pix2Pix), and three-dimensional Cycle-Consistent Generative Adversarial Network (3D-CycleGAN) p rimarily using Dice Similarity Coefficient and sensitivity. Radiomic features were extracted from predicted MTV (Pred_MTV) and GT_MTV to construct Cox proportional hazards models for event-free survival, evaluated by concordance index (C-index). Robustness was further assessed in an internal temporal validation cohort from 2022 with different scanners (n = 135) using planning CT as the sole input.
RESULTS: MedNext-Insight achieved the highest MTV delineation Dice Similarity Coefficient (Mean ± SD, 0.808 ± 0.110 vs 0.740-0.782; all P < .05) with improved sensitivity. After excluding 12 patients with baseline distant metastasis from event-free survival analysis, Pred_MTV-derived radiomics showed strong reproducibility (median intraclass correlation coefficient, 0.816) and comparable prognostic performance to GT_MTV (C-index [95% CI], 0.712 [0.516-0.899] vs 0.744 [0.601-0.884]; P = .730). MTV-based radiomics outperformed primary gross tumor volume-derived features, particularly when combined with clinical variables (C-index, 0.809 [0.678-0.919]). In the internal temporal validation cohort, CT-only Pred_MTV maintained stable segmentation accuracy and prognostic discrimination (log-rank test, P < .05).
CONCLUSION: MedNext-Insight enables accurate PET-free MTV delineation on routine CT with prognostic value, supporting a resource-efficient approach for risk stratification and informing potential future biology-guided adaptive radiation therapy in NPC.
METHODS AND MATERIALS: A retrospective cohort of 392 patients with NPC undergoing pre-radiotherapy 2-deoxy-2-[fluorine-18]fluoro-D-glucose PET/CT in 2021 was enrolled and randomly divided into training (n = 314, including 63 for validation) and test (n = 78) cohorts. Ground-truth MTV (GT_MTV) was generated from PET-registered CT using standardized uptake value SUV > 2.5 within the primary gross tumor volume. A 7-layer MedNext-Insight model with dual-window CT inputs and Dice-Focal loss was trained to predict MTV using CT alone. Segmentation performance was compared with no-new-U-Net version 2 (nnUNetV2), Conditional Generative Adversarial Network for Image-to-Image Translation (Pix2Pix), and three-dimensional Cycle-Consistent Generative Adversarial Network (3D-CycleGAN) p rimarily using Dice Similarity Coefficient and sensitivity. Radiomic features were extracted from predicted MTV (Pred_MTV) and GT_MTV to construct Cox proportional hazards models for event-free survival, evaluated by concordance index (C-index). Robustness was further assessed in an internal temporal validation cohort from 2022 with different scanners (n = 135) using planning CT as the sole input.
RESULTS: MedNext-Insight achieved the highest MTV delineation Dice Similarity Coefficient (Mean ± SD, 0.808 ± 0.110 vs 0.740-0.782; all P < .05) with improved sensitivity. After excluding 12 patients with baseline distant metastasis from event-free survival analysis, Pred_MTV-derived radiomics showed strong reproducibility (median intraclass correlation coefficient, 0.816) and comparable prognostic performance to GT_MTV (C-index [95% CI], 0.712 [0.516-0.899] vs 0.744 [0.601-0.884]; P = .730). MTV-based radiomics outperformed primary gross tumor volume-derived features, particularly when combined with clinical variables (C-index, 0.809 [0.678-0.919]). In the internal temporal validation cohort, CT-only Pred_MTV maintained stable segmentation accuracy and prognostic discrimination (log-rank test, P < .05).
CONCLUSION: MedNext-Insight enables accurate PET-free MTV delineation on routine CT with prognostic value, supporting a resource-efficient approach for risk stratification and informing potential future biology-guided adaptive radiation therapy in NPC.