Identification of paraglottic space invasion in enhanced CT scans of hypopharyngeal cancer by 3D super-resolution reconstruction technology and deep learning
10.3760/cma.j.cn115330-20241230-00717
- VernacularTitle:利用3D超分辨率重建技术和深度学习识别下咽癌增强CT声门旁间隙受侵
- Author:
Wenlun WANG
1
;
Zhiwei LIU
1
;
Jing′ao LI
1
;
Chenyang XU
1
;
Dongmin WEI
1
;
Ye QIAN
1
;
Wenming LI
1
;
Dapeng LEI
1
Author Information
1. 山东大学齐鲁医院耳鼻咽喉科 国家卫生健康委员会耳鼻咽喉科学重点实验室（山东大学），济南 250012
- Publication Type:Journal Article
- Keywords:
Hypopharyngeal neoplasms;
Paraglottic space invasion;
Super-resolution reconstruction;
Deep learning
- From:
Chinese Journal of Otorhinolaryngology Head and Neck Surgery
2025;60(10):1232-1242
- CountryChina
- Language:Chinese
-
Abstract:
Objective:To develop a deep learning model based on 3D super-resolution reconstruction technology and to analyze its feasibility and effectiveness in predicting paraglottic space invasion in hypopharyngeal cancer.Methods:A retrospective study was conducted involving 382 patients with hypopharyngeal squamous cell carcinoma treated at Qilu Hospital of Shandong University between January 2014 and December 2020. The cohort included 364 males and 18 females, with a mean age of 62±7 years. Patients were divided into a training set ( n=300) and a test set ( n=82) based on enrollment time. A generative adversarial network was used to perform 3D super-resolution reconstruction on contrast-enhanced CT images, improving spatial resolution by 16 times. A 2.5D deep learning strategy was employed to construct Resnet-NR and Resnet-SR models based on conventional and super-resolution images, respectively, to predict whether the paraglottic space was invaded. Model performance was evaluated using receiver operating characteristic (ROC) curves and area under the curve (AUC). A multi-reader multi-case study was conducted to assess the impact of the artificial intelligence (AI) model on clinicians′ diagnostic capabilities. Results:The super-resolution model Resnet-SR achieved the highest accuracy in both the training set (AUC=0.87, 95% CI: 0.84-0.90) and the test set (AUC=0.88, 95% CI: 0.81-0.96), significantly outperforming traditional clinical indicators (T stage, N stage, tumor diameter, and pathological differentiation degree) (AUC range: 0.55-0.70, all P<0.05). In comparison, the conventional-resolution model Resnet-NR achieved AUCs of 0.81 (95% CI: 0.77-0.84, P=0.005) and 0.80 (95% CI: 0.71-0.89, P=0.184) in the training and test sets, respectively. Using Resnet-SR to assist clinical decision-making improved the diagnostic accuracy of junior physicians (AUC=0.793 without AI assistance vs. AUC=0.871 with AI assistance, P=0.012) and significantly reduced diagnosis time for clinicians of all experience levels (86.5 s without AI assistance vs. 82.5 s with AI assistance, t=2.01, P=0.032). Conclusion:This study successfully develops a deep learning model based on 3D super-resolution reconstruction technology, which can assist in preoperative prediction of paraglottic space invasion in hypopharyngeal cancer. The AI-assisted tool improves diagnostic accuracy for junior physicians and enhances diagnostic efficiency for clinicians across all experience levels.
