Robust Unsupervised Super-Resolution of Infant MRI via Dual-Modal Deep Image Prior

Area of Science:

• Medical Imaging
• Neuroscience
• Computer Vision

Background:

• Infant brain development studies rely on Magnetic Resonance Imaging (MRI), but acquisition challenges like long scan times and motion artifacts limit data quality.
• Existing super-resolution (SR) methods often require paired low-resolution (LR) and high-resolution (HR) images, which are impractical for infant MRI.
• Deep Image Prior (DIP) offers unsupervised single-image SR but struggles with automated early stopping criteria.

Purpose of the Study:

• To develop an unsupervised super-resolution technique for enhancing infant MRI quality without additional imaging burden.
• To address the challenge of automated early stopping in Deep Image Prior (DIP) for single-image SR.
• To leverage multi-modal MRI data (T1-weighted and T2-weighted) for improved image reconstruction.

Main Methods:

• A novel dual-modal Deep Image Prior (dmDIP) framework was designed, integrating information from both T1-weighted and T2-weighted MRI scans.
• The method constrains the low-frequency k-space of the super-resolved image to match that of the input low-resolution image.
• The unsupervised approach optimizes the image reconstruction using only the low-resolution input, eliminating the need for paired high-resolution data.

Main Results:

• The dmDIP model successfully enhanced the image quality of infant MRI scans.
• The proposed method demonstrated substantially reduced sensitivity to the critical early stopping parameter in DIP training.
• Evaluation on infant MRI data from birth to one year confirmed the effectiveness of the dual-modal approach.

Conclusions:

• Dual-modal Deep Image Prior (dmDIP) provides an effective solution for unsupervised super-resolution of infant MRI.
• The technique improves image quality and robustness to training parameters, facilitating automated processing.
• dmDIP holds promise for advancing the study of early brain development using high-quality MRI data.