Synthesizing Contrast-Enhanced MR Images from Noncontrast MR Images Using Deep Learning

Area of Science:

• Artificial Intelligence
• Medical Imaging
• Deep Learning

Background:

• Gadolinium-based contrast agents raise toxicity concerns, necessitating alternative imaging methods.
• Deep learning offers a promising approach to develop novel imaging techniques.
• Noncontrast multiparametric MRI is a safer alternative for brain tumor imaging.

Purpose of the Study:

• To synthesize virtual gadolinium contrast-enhanced T1-weighted MR images from noncontrast multiparametric MR images in patients with primary brain tumors.
• To evaluate the efficacy of deep learning in generating realistic contrast-enhanced images.
• To assess the potential of reducing gadolinium contrast agent usage.

Main Methods:

• A deep learning network (T1c-ET) was developed using the Brain Tumor Segmentation Challenge 2019 dataset (335 subjects for training, 125 for testing).
• The network was trained to simultaneously synthesize virtual contrast-enhanced T1-weighted (vT1c) images and segment enhancing tumor portions.
• Synthesized vT1c images were independently scored by three neuroradiologists for image quality and contrast enhancement.

Main Results:

• The synthesized vT1c images achieved high quality scores (SSIM: 0.91, PSNR: 64.35, NMSE: 0.03).
• The model accurately predicted contrast enhancement in 88.8% of cases.
• Moderate interobserver agreement (Fleiss kappa = 0.61) was observed for contrast enhancement prediction.

Conclusions:

• A novel deep learning architecture was developed to synthesize virtual contrast enhancement from noncontrast brain MRI.
• The study demonstrates the potential of deep learning to reduce reliance on gadolinium contrast agents for primary brain tumor evaluation.
• This AI-driven approach offers a safer and potentially more accessible method for brain tumor imaging.