ECLARE: Efficient cross-planar learning for anisotropic resolution enhancement

TL;DR

ECLARE tackles the challenge of super-resolving anisotropic multi-slice 2D MR volumes by explicitly modeling the slice selection profile and through-plane sampling, then training a self-supervised SR network on HF in-plane patches degraded with a learned PSF. It combines ESPRESO-based slice-profile estimation, FOV-aware interpolation, and a WDSR-based network capable of non-integer upsampling, enabling SR without external data. Across two datasets and multiple resolutions, ECLARE consistently improves PSNR, SSIM, and downstream segmentation metrics, while avoiding domain-shift issues inherent to pretrained models. The approach is efficient, openly available, and broadly applicable to MR contrasts and pathologies, though computation time per volume remains a limitation for real-time use.

Abstract

In clinical imaging, magnetic resonance (MR) image volumes are often acquired as stacks of 2D slices with decreased scan times, improved signal-to-noise ratio, and image contrasts unique to 2D MR pulse sequences. While this is sufficient for clinical evaluation, automated algorithms designed for 3D analysis perform poorly on multi-slice 2D MR volumes, especially those with thick slices and gaps between slices. Super-resolution (SR) methods aim to address this problem, but previous methods do not address all of the following: slice profile shape estimation, slice gap, domain shift, and non-integer or arbitrary upsampling factors. In this paper, we propose ECLARE (Efficient Cross-planar Learning for Anisotropic Resolution Enhancement), a self-SR method that addresses each of these factors. ECLARE uses a slice profile estimated from the multi-slice 2D MR volume, trains a network to learn the mapping from low-resolution to high-resolution in-plane patches from the same volume, and performs SR with anti-aliasing. We compared ECLARE to cubic B-spline interpolation, SMORE, and other contemporary SR methods. We used realistic and representative simulations so that quantitative performance against ground truth can be computed, and ECLARE outperformed all other methods in both signal recovery and downstream tasks. Importantly, as ECLARE does not use external training data it cannot suffer from domain shift between training and testing. Our code is open-source and available at https://www.github.com/sremedios/eclare.

Figures (9)

• Figure 1: A flowchart of our proposed method. The anisotropic input volume is fed independently into each of the three steps. First, in (A) (Sec. \ref{['sec:approx_imaging']}), we estimate the slice excitation profile with ESPRESO han2023espreso. Second, in (B) (Sec. \ref{['sec:eclare_training']}), we extract HR in-plane 2D patches and use the PSF estimated from (A) to create paired training data. This training data is used to train the network $f_\theta$ with supervised learning. Third, in (C) (Sec. \ref{['sec:apply_model']}), we extract LR through-plane 2D slices and super-resolve them with the trained network $f_\theta$ from (B).The super-resolved slices are stacked and averaged, yielding the super-resolved output volume.
• Figure 2: A 1D demonstration of the differences between interpolation implementations. The input has 6 pixels and its FOV center is located at 2.5, indicated by the red dashed line. Its sample spacing is 1, indicated by the purple arrow. Its extents occur at -0.5 and 5.5, indicated by the blue dashed line. In this demonstration, the desired sample spacing for upsampling is 0.7, and the desired sample spacing for downsampling is 1.429. The implementation scipy.zoomscipy_zoom maintains the FOV center and boundary sample locations, but changes the sample spacing to 0.625 and 1.667 for upsampling and downsampling, respectively. The implementation torch.nn.functional.interpolatepytorch anchors the FOV extent to start at 0.5, shifting the center from the desired FOV center. Our proposed method (Sec. \ref{['sec:interp']}) maintains both the sample spacing and the FOV center.
• Figure 3: A diagram of our proposed architecture, based on WDSR yu2018wide. The low resolution image $y$ is first processed by a convolution to produce a feature map with 256 channels. It then passes through several WDSR blocks. A final convolution is done to produce the correct number of channels for pixel shuffle. Finally, the result is downsampled and added to an upsampled $y$ to yield the output. In our implementation, we use 16 WDSR blocks.
• Figure 4: A $5\|1.5$ sagittal slice from a representative subject from the OASIS-3 dataset. The first row shows the image domain with PSNR and SSIM for the entire image volume overlaid. The second row shows the simplified SLANT labels, with an overlay showing the mean cDSC across all labels for the entire image volume.
• Figure 5: Quantitative results (PSNR on the left, SSIM on the right) on the OASIS-3 dataset across resolutions.