Deep learning-based non-contrast MRI model for nasopharyngeal carcinoma diagnosis: an end-to-end gadolinium-free solution.
Authors
Affiliations (20)
Affiliations (20)
- ENT Institute and Department of Otorhinolaryngology, Eye & ENT Hospital, Fudan University, Shanghai, China.
- Yale Biomedical Imaging Institute and Department of Radiology and Biomedical Imaging, Yale University, New Haven, CT, USA.
- Department of Radiology, Eye & ENT Hospital, Fudan University, Shanghai, China.
- Department of Electrical and Computer Engineering, Boston University, Boston, MA, USA.
- Gordon Center for Medical Imaging, Massachusetts General Hospital and Harvard Medical School, Boston, MA, USA.
- Department of Radiation Oncology, Shanghai Proton and Heavy Ion Center, Fudan University Cancer Hospital, Shanghai, China.
- Shanghai Key Laboratory of Radiation Oncology, Shanghai, China.
- Shanghai Engineering Research Center of Proton and Heavy Ion Radiation Therapy, Shanghai, China.
- Department of Radiology, Shanghai Sixth People's Hospital, Shanghai, China.
- Department of Surgery, Brigham and Women's Hospital, Harvard Medical School, Boston, MA, USA.
- Department of Radiology, Nanjing Drum Tower Hospital Group Suqian Hospital, Suqian, China.
- Department of Radiology, Jiangsu Province Suqian Hospital, Suqian, China.
- Department of Biomedical Informatics and Data Science, Yale University, New Haven, CT, USA.
- Medical Artificial Intelligence Laboratory, Westlake University, Hangzhou, China.
- ENT Institute and Department of Otorhinolaryngology, Eye & ENT Hospital, Fudan University, Shanghai, China. [email protected].
- ENT Institute and Department of Otorhinolaryngology, Eye & ENT Hospital, Fudan University, Shanghai, China. [email protected].
- Research Units of New Technologies of Endoscopic Surgery in Skull Base Tumor, Chinese Academy of Medical Sciences, Shanghai, China. [email protected].
- Yale Biomedical Imaging Institute and Department of Radiology and Biomedical Imaging, Yale University, New Haven, CT, USA. [email protected].
- Department of Biomedical Informatics and Data Science, Yale University, New Haven, CT, USA. [email protected].
- Imaging and Data Sciences Platforms, Broad Institute of MIT and Harvard, Boston, MA, USA. [email protected].
Abstract
Nasopharyngeal carcinoma (NPC) diagnosis and routine follow-up for recurrence typically rely on contrast-enhanced MRI. This study introduces a deep learning model for diagnosing NPC using only non-contrast MRI, reducing the need for gadolinium-based contrast agents (GBCA). This approach helps avoid safety concerns related to GBCA deposition, while also shortening scan times and reducing costs. In this study, we propose an innovative deep learning model for NPC diagnosis using only non-contrast MRI, reducing the need for gadolinium-based contrast agents (GBCA). This approach not only mitigates potential safety concerns associated with residual GBCA deposition but also reduces scan time and examination costs. The study consisted of three phases. Firstly, a novel knowledge distilled modality fusion model is developed using a cohort of 854 cases and tested its performance on an internal set (257 cases, AUC = 0.95) and an independent external set (277 cases, AUC = 0.86). Secondly, the proposed method was compared with: (1) Non-contrast MRI without model improvement (Baseline 1) and (2) current virtual-contrast enhancement-based NPC diagnosis using three state-of-the-art methods (Baselines 2-4). The proposed model consistently outperformed Baselines 1 on both internal dataset (AUC: 0.95 vs. 0.93) and external test set (AUC: 0.86 vs. 0.82). Additionally, it surpassed Baselines 2-4, achieving performance gains of 6.7%, 69.8%, and 28.6% in AUC, over three state-of-the-art methods. Thirdly, the effectiveness of this model was evaluated through a fully crossed multi-reader, multi-case study involving 13 readers from 6 hospitals. The results showed that with AI assistance, readers could diagnose NPC using only non-contrast MRI, achieving results that were not inferior to contrast-enhanced imaging (AUC: 0.90 vs. 0.93, p<0.01). In conclusion, this study demonstrated the model's potential as a safe, cost-effective, and GBCA-free option for NPC diagnosis in clinical practice.