Multi-institutional Investigation of Model Generalizability for Virtual Contrast-Enhanced MRI Synthesis

The purpose of this study is to investigate the model generalizability using multi-institutional data for virtual contrast-enhanced MRI (VCE-MRI) synthesis. This study presented a retrospective analysis of contrast-free T1-weighted (T1w), T2-weighted (T2w), and gadolinium-based contrast-enhanced T1w MRI (CE-MRI) images of 231 NPC patients enrolled from four institutions. Data from three of the participating institutions were employed to generate a training and an internal testing set, while data from the remaining institution was employed as an independent external testing set. The multi-institutional data were trained separately (single-institutional model) and jointly (joint-institutional model) and tested using the internal and external sets. The synthetic VCE-MRI was quantitatively evaluated using MAE and SSIM. In addition, visual qualitative evaluation was performed to assess the quality of synthetic VCE-MRI compared to the ground-truth CE-MRI. Quantitative analyses showed that the joint-institutional models outperformed single-institutional models in both internal and external testing sets, and demonstrated high model generalizability, yielding top-ranked MAE, and SSIM of 71.69 ± 21.09 and 0.81 ± 0.04 respectively on the external testing set. Qualitative evaluation indicated that the joint-institutional model gave a closer visual approximation between the synthetic VCE-MRI and ground-truth CE-MRI on the external testing set, compared with single-institutional models. The model generalizability for VCE-MRI synthesis was enhanced, both quantitatively and qualitatively, when data from more institutions was involved during model development.