Can diffusion-based generated magnetic resonance images predict glioma methylation accurately?

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Radiology

Background:

• Accelerated magnetic resonance imaging (MRI) reconstruction is vital for medical diagnostics, especially for brain tumors.
• Existing deep learning methods struggle to preserve diagnostic features across contrasts during accelerated scans.
• There is a clinical need for methods that maintain pathological details while reducing scan times.

Purpose of the Study:

• To develop a generative model for reconstructing high-quality MRI sequences from accelerated T1-weighted scans.
• To preserve pathological features critical for glioma diagnosis and MGMT prediction.
• To enable faster MRI acquisition without compromising diagnostic accuracy.

Main Methods:

• A retrospective study using the BraTS 2021 dataset (n=1,480).
• Development and application of a diffusion-based generative model for T1CE, T2, and FLAIR sequences.
• Simulation of 4-fold and 32-fold acceleration by K-space degradation.

Main Results:

• High-fidelity reconstruction with excellent SSIM and PSNR values for 4-fold acceleration.
• Radiologist assessments showed no significant differences between original and reconstructed images.
• Consistent MGMT methylation classification and stable AUC values for prediction across accelerations.

Conclusions:

• The diffusion-based model effectively accelerates MRI acquisition while preserving pathological details.
• The model shows promise for enhanced diagnostic capabilities in neuro-oncology.
• Future clinical evaluation is planned to assess procedural efficiency and patient comfort.