ECLARE: efficient cross-planar learning for anisotropic resolution enhancement

Area of Science:

• Medical Imaging
• Image Processing
• Computational Anatomy

Background:

• Clinical magnetic resonance (MR) imaging often uses 2D slice acquisition for efficiency, leading to anisotropic volumes with thick slices and gaps.
• Existing 3D analysis algorithms struggle with these anisotropic MR volumes, impacting automated analysis.
• Current super-resolution (SR) methods fail to address critical factors like slice profile, gaps, domain shift, and arbitrary upsampling.

Purpose of the Study:

• To introduce ECLARE (Efficient Cross-planar Learning for Anisotropic Resolution Enhancement), a novel self-supervised SR method.
• To address limitations of previous SR techniques by incorporating slice profile estimation, gap handling, and FOV-aware resampling.
• To improve the performance of automated 3D analysis algorithms on anisotropic MR image volumes.

Main Methods:

• ECLARE estimates slice profiles directly from the 2D MR volume.
• It employs a self-supervised approach, training a network to map low-resolution to high-resolution in-plane patches from the same volume.
• The method incorporates antialiasing and respects the field of view (FOV) during resampling, validated on T1-w and T2-w FLAIR datasets.

Main Results:

• ECLARE significantly outperforms contemporary SR methods and B-spline interpolation in mean Peak Signal-to-Noise Ratio (PSNR) and Structural Similarity Index Measure (SSIM) for images with up to 5mm slice thickness and 1.5mm gaps.
• Performance is comparable or superior to other methods in key brain regions like ventricles, caudate, and white matter.
• Consistent results were observed across both healthy T1-w and multiple sclerosis (MS) T2-w FLAIR datasets.

Conclusions:

• ECLARE's integrated approach (slice profile estimation, FOV-aware resampling, self-SR) enables robust super-resolution of anisotropic MR images without external training data.
• The method demonstrates significant potential for enhancing medical image analysis by improving resolution and consistency.
• Future work will explore ECLARE's applicability across different organs, species, modalities, and resolutions, with open-source code available.