Model Generalizability Investigation for GFCE-MRI Synthesis in NPC Radiotherapy Using Multi-Institutional...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Radiology

Background:

• Gadolinium-based contrast agents (GBCAs) pose safety risks.
• Deep learning offers gadolinium-free contrast-enhanced MRI (GFCE-MRI) synthesis as an alternative.
• Inter-institutional heterogeneity and data scarcity challenge GFCE-MRI model generalizability.

Purpose of the Study:

• To investigate the generalizability of GFCE-MRI models across multiple institutions.
• To evaluate the impact of data normalization methods on model performance.
• To identify optimal strategies for robust GFCE-MRI synthesis.

Main Methods:

• Utilized data from seven institutions for GFCE-MRI synthesis in nasopharyngeal carcinoma patients.
• Applied five popular data normalization approaches to address MRI heterogeneity.
• Trained uni-institution and tri-institution models using T1-weighted and T2-weighted MRI.
• Assessed model generalizability using external validation cohorts from four institutions.

Main Results:

• Uni-institution models showed significant performance degradation on external cohorts.
• Multi-institutional training improved model generalizability compared to uni-institution models.
• Z-Score normalization, when applied to multi-institutional data, yielded the best generalizability.

Conclusions:

• GFCE-MRI models trained on single institutions lack generalizability.
• Multi-institutional data and appropriate normalization are crucial for robust GFCE-MRI synthesis.
• Z-Score normalization is a promising approach for enhancing GFCE-MRI model generalizability.