Contrast-free identification of glioma blood-brain barrier status via generative diffusion AI and non-contrast MRI.
Authors
Affiliations (13)
Affiliations (13)
- School of Biomedical Engineering, Southern Medical University, Guangzhou, China.
- Guangdong Provincial Key Laboratory of Medical Image Processing, Guangzhou, China.
- Guangdong Province Engineering Laboratory for Medical Imaging and Diagnostic Technology, Guangzhou, China.
- Department of Biostatistics, School of Global Public Health, New York University, New York, USA.
- Department of Radiology, Zhujiang Hospital, Southern Medical University, Guangzhou, China.
- School of Life Science and Technology, Xi'an Jiaotong University, Xi'an, China.
- Department of Medical Imaging Center, Nanfang Hospital, Southern Medical University, Guangzhou, China. [email protected].
- School of Biomedical Engineering, Southern Medical University, Guangzhou, China. [email protected].
- Guangdong Provincial Key Laboratory of Medical Image Processing, Guangzhou, China. [email protected].
- Guangdong Province Engineering Laboratory for Medical Imaging and Diagnostic Technology, Guangzhou, China. [email protected].
- School of Biomedical Engineering, Southern Medical University, Guangzhou, China. [email protected].
- Guangdong Provincial Key Laboratory of Medical Image Processing, Guangzhou, China. [email protected].
- Guangdong Province Engineering Laboratory for Medical Imaging and Diagnostic Technology, Guangzhou, China. [email protected].
Abstract
Non-contrast MRI, routinely used for the preoperative diagnosis of glioma tumors and establishing treatment strategies, provides the potential for assessing blood-brain barrier (BBB) status without using gadolinium-based contrast agents (GBCA) which could cause adverse events. Additionally, generative artificial intelligence (AI) models enable the synthesis of contrast-enhanced images from non-contrast images. Despite this potential, the heterogeneity of GBCA-induced features in tumor areas and error accumulation from inaccurate synthesis largely limit the efficacy of conventional generative models. To address these limitations, we introduce a contrast-free BBB status identification model (CBSI) that can identify BBB status with high accuracy using non-contrast MR images and generative diffusion AI networks. Trained and validated on a multi-center dataset of 1,535 patients, CBSI achieves an area under the curve (AUC) of 81.31%, surpassing the performance of the model using only non-contrast MR (AUC = 72.76%) and demonstrating comparable performance to the T1Gd MR model (AUC = 88.68%) in an external test set. Furthermore, validation on two public datasets (BraTS-Africa and BraTS-GLI) supports the generalizability of CBSI in BBB status identification. Notably, with accurate BBB status of synthetic T1Gd, the performance of glioma segmentation and grading is improved significantly compared to existing methods. Generalizability analysis indicates that CBSI can facilitate BBB status identification using synthetic T1Gd findings, avoiding GBCA adverse effects and streamlining clinical workflows.