Computer methods and programs in biomedicine · 2025
A novel multimodal adaptive delineation model for primary tumors and lymph node metastases in multi-center nasopharyngeal carcinoma radiotherapy.
СтатьяМногоцентровое исследованиеНазвание 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.
Sun Hongfei, Duan Jie, Zhu Jiarui, Li Zhihui, Liu Yufen, Liu Changhao, Li Jie, Shi Zihan, Li Ningning, Gong Jie, Li Xiaokai, Wang Zhongfei, Li Dong, Shi Mei, Zhao Lina
Аннотация
PURPOSE: Nasopharyngeal carcinoma (NPC) features uncertain and complex gross tumor volume (GTV) distributions in terms of location, size, and shape. A novel deep learning model with adaptability was proposed to improve the accuracy of automatic GTV delineation for NPC primary and metastatic lesions.
METHODS AND MATERIALS: This study comprised 529 retrospective cases and 4 prospective cases from multiple centers, all with CT and MRI modalities. GTV adaptive delineation was achieved via a conditional denoising diffusion model (DDPM) with "inter-modal and intra-modal" attention aware mechanism. During one training epoch, the inter-modal aware mechanism linked the frequency of potentially effective features at identical coordinates across multimodal images to tumor locations. The model progressively focused on high-frequency GTV-related features. Across multiple trainings, the intra-modal aware mechanism established the repeatability of feature extraction within each modality at identical coordinates, enhancing stable feature extraction. By leveraging both mechanisms, the model dynamically calibrated fusion weights for multimodal features, ascertaining each feature's significance in GTV identification based on its positional frequency. The GTV delineation accuracy was assessed using Dice Similarity Coefficient (DSC) ( %), 95 % Hausdorff Distance (HD95 %) (mm), and Mean Surface Distance (MSD) (mm) metrics.
RESULTS: For the internal test set, the adaptive delineation model yielded mean (SD) results of 81.36 ± 2.14 % for DSC, 4.30 ± 2.14 mm for HD95 %, and 3.70 ± 1.84 mm for MSD. The external test set showed corresponding values of 77.43 ± 3.41 %, 6.07 ± 3.10 mm, and 4.31 ± 2.64 mm. Paired T-tests confirmed statistically significant differences between our model and the current SOTA models for automatic GTV delineation. The adaptive delineation model attained DSC accuracies of 83.45 ± 1.75 % for primary tumor and 73.43 ± 5.32 % for metastatic lesions, while achieving HD95 precisions of 4.28 ± 2.08 mm and 7.33 ± 3.45 mm, respectively. The MSD measurements were 2.73 ± 1.70 mm and 5.90 ± 3.18 mm, respectively. These results highlighted better delineation accuracy for primary tumors. In prospective dosimetry validation, the average dose difference within primary tumors, comparing automatically and manually delineated GTVs, was 0.36 Gy-less than the 0.52 Gy difference in lymph node metastases. This aligned with retrospective validation patterns.
CONCLUSION: The novel deep learning model demonstrated high accuracy and stability GTV automatic delineation for NPC cases, indicating its potential for clinical application in NPC radiotherapy at different centers.
METHODS AND MATERIALS: This study comprised 529 retrospective cases and 4 prospective cases from multiple centers, all with CT and MRI modalities. GTV adaptive delineation was achieved via a conditional denoising diffusion model (DDPM) with "inter-modal and intra-modal" attention aware mechanism. During one training epoch, the inter-modal aware mechanism linked the frequency of potentially effective features at identical coordinates across multimodal images to tumor locations. The model progressively focused on high-frequency GTV-related features. Across multiple trainings, the intra-modal aware mechanism established the repeatability of feature extraction within each modality at identical coordinates, enhancing stable feature extraction. By leveraging both mechanisms, the model dynamically calibrated fusion weights for multimodal features, ascertaining each feature's significance in GTV identification based on its positional frequency. The GTV delineation accuracy was assessed using Dice Similarity Coefficient (DSC) ( %), 95 % Hausdorff Distance (HD95 %) (mm), and Mean Surface Distance (MSD) (mm) metrics.
RESULTS: For the internal test set, the adaptive delineation model yielded mean (SD) results of 81.36 ± 2.14 % for DSC, 4.30 ± 2.14 mm for HD95 %, and 3.70 ± 1.84 mm for MSD. The external test set showed corresponding values of 77.43 ± 3.41 %, 6.07 ± 3.10 mm, and 4.31 ± 2.64 mm. Paired T-tests confirmed statistically significant differences between our model and the current SOTA models for automatic GTV delineation. The adaptive delineation model attained DSC accuracies of 83.45 ± 1.75 % for primary tumor and 73.43 ± 5.32 % for metastatic lesions, while achieving HD95 precisions of 4.28 ± 2.08 mm and 7.33 ± 3.45 mm, respectively. The MSD measurements were 2.73 ± 1.70 mm and 5.90 ± 3.18 mm, respectively. These results highlighted better delineation accuracy for primary tumors. In prospective dosimetry validation, the average dose difference within primary tumors, comparing automatically and manually delineated GTVs, was 0.36 Gy-less than the 0.52 Gy difference in lymph node metastases. This aligned with retrospective validation patterns.
CONCLUSION: The novel deep learning model demonstrated high accuracy and stability GTV automatic delineation for NPC cases, indicating its potential for clinical application in NPC radiotherapy at different centers.