Zero-shot arbitrary-scale super resolution in susceptibility-weighted imaging for cerebral microbleed analysis

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Neuroscience

Background:

• Susceptibility Weighted Imaging (SWI) is crucial for identifying cerebral microbleeds (CMBs), indicators of vascular issues and neurodegeneration.
• Clinical SWI protocols often sacrifice resolution for speed, compromising the detection of subtle through-plane details.
• Existing deep learning super-resolution (SR) methods require extensive paired high-resolution (HR) and low-resolution (LR) data, which is challenging to obtain in medical contexts.

Purpose of the Study:

• To introduce MagNeRF, a novel zero-shot, single-subject SR framework for arbitrary-scale reconstruction of SWI.
• To enable high-resolution SWI reconstruction from a single low-resolution volume without external paired training data.
• To improve the detection and characterization of cerebral microbleeds (CMBs) in SWI.

Main Methods:

• MagNeRF employs a dilated patch-based sampling strategy to enhance spatial context and local detail recovery.
• A spherical sampling strategy is utilized to effectively capture the radial gradient decay characteristic of SWI magnetic susceptibility signals.
• Adaptive loss functions, including adaptive multi-scale structural similarity and adaptive mean squared error, are implemented to prioritize perceptual fidelity and structural integrity.

Main Results:

• MagNeRF demonstrated superior performance compared to state-of-the-art methods on SWI datasets for CMB detection, yielding high visual fidelity and preserving diagnostically relevant structures.
• The framework's robustness was confirmed across diverse MRI contrasts, including T1-weighted and real-world LR T2*-weighted datasets.
• Downstream CMB lesion segmentation using MagNeRF-reconstructed HR images achieved performance comparable to original HR data, highlighting its clinical relevance.

Conclusions:

• MagNeRF shows significant promise for preserving clinically relevant microbleed features in SWI, offering a powerful tool for enhancing diagnostic accuracy.
• The ability to reconstruct HR SWI from LR inputs can reduce patient scanning time and burden.
• MagNeRF has the potential to broaden the clinical applicability of SWI in diagnosing and monitoring neurovascular and neurodegenerative conditions.