NEIM-03 A GRADUAL INTEGRATION MODEL BASED ON MRI TO PREDICT MOLECULAR PROFILES IN PATIENTS WITH GLIOMA

Abstract

Abstract Glioma is the most common primary brain tumor. Molecular profiles, including IDH mutation, 1p/19q co-deletion, and MGMT promoter methylation, are highly correlated with the prognosis and clinical decision-making of glioma. To predict the molecular profiles from rough segmentation MRI, we developed and validated a gradual integration model. We proposed a multi-task pseudo-3D model based on rough segmented multiparametric MRI to predict molecular profiles. In this model, convolution along the depth axis was performed between convolution blocks. A total of 750 patients with glioma were retrospectively enrolled from Sun Yatsen University Cancer Center(n=576) and The Cancer Imaging Archive(n=167) to validate the performance of the framework. The model was developed and validated on the local dataset and tested on an independent external (TCIA) test set. For IDH, 1p/19q, and MGMT predictions on the TCIA set, the gradual fusion model achieves AUCs of 0.87, 0.80, and 0.71, respectively, outperforming both input or output integration models and 2D models. To conclude, our gradual fusion model demonstrated the potential to predict molecular profiles based on pre-treatment MRI. Gradual integration may be a reliable method to design 3D medical imaging models.