Super-Resolution MR Spectroscopic Imaging via Diffusion Models for Tumor Metabolism Mapping

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Neuro-oncology

Background:

• High-resolution magnetic resonance spectroscopic imaging (MRSI) is vital for glioma patient care, but low signal-to-noise ratio limits its spatial resolution.
• Current low-resolution MRSI hinders the accurate assessment of tumor heterogeneity and margins.

Purpose of the Study:

• To develop a novel deep learning framework for super-resolution reconstruction of MRSI data.
• To enhance the visualization of tumor metabolism and boundaries in mutant isocitrate dehydrogenase (IDH) gliomas.

Main Methods:

• A conditional denoising diffusion probabilistic model with a Self-Attention UNet backbone was employed for MRSI super-resolution.
• The model progressively reconstructs high-fidelity metabolite maps from low-resolution inputs through a learned reverse diffusion process.
• The framework was validated using simulated data and in vivo MRSI from healthy volunteers and glioma patients.

Main Results:

• The proposed method demonstrated superior performance in quantitative metrics (SSIM, PSNR, LPIPS) across various upsampling factors (2x, 4x, 8x) on simulated data.
• Statistically significant improvements in LPIPS were observed compared to baseline methods (p < 0.01).
• In vivo validation showed accurate reconstruction of small lesions, preservation of textural details, and enhanced tumor boundary delineation, revealing previously unseen metabolic heterogeneity.

Conclusions:

• Diffusion-based deep learning models offer a promising approach for noninvasive, high-resolution metabolic imaging in glioma.
• The developed framework can significantly improve the clinical utility of MRSI for diagnosing and managing gliomas.
• This technology has potential applications in other neurological disorders requiring detailed metabolic assessment.