SIMPLE: Simultaneous Multi-Plane Self-Supervised Learning for Isotropic MRI Restoration from Anisotropic Data

TL;DR

The study tackles the challenge of isotropic MRI restoration from anisotropic data without relying on simulated downsampling or HR training pairs. It proposes SIMPLE, a simultaneous multi-plane self-supervised framework that combines a 3D U-Net generator with plane-specific 2D discriminators, incorporating a pre-training step on single-plane SR via ATME and a multi-plane adversarial–consistency loss formulation. Evaluations on brain (OASIS) and abdomen (Crohn's) datasets show that SIMPLE achieves superior isotropic quality across planes, as evidenced by lower KID/FID/IS metrics and favorable radiologist assessments, as well as improved Fourier-domain properties. The approach enables more accurate volumetric analyses and 3D reconstructions, with strong potential for clinical adoption and extension to other contrasts and acquisition methods. The method is mathematically framed by relations such as $V_{ ext{An-Iso}}$, $V'_{ ext{Iso}} = L(V_{ ext{An-Iso}})$, and the isotropic reconstruction $\uhat{V}_{ ext{Iso}} = GM(V'_{ ext{Iso}})$ under cross-plane constraints $S_{ ext{HR}_{ ext{Cor}}} = t$ and similar formulations, ensuring consistent high-resolution outputs across all three planes.

Abstract

Magnetic resonance imaging (MRI) is crucial in diagnosing various abdominal conditions and anomalies. Traditional MRI scans often yield anisotropic data due to technical constraints, resulting in varying resolutions across spatial dimensions, which limits diagnostic accuracy and volumetric analysis. Super-resolution (SR) techniques aim to address these limitations by reconstructing isotropic high-resolution images from anisotropic data. However, current SR methods often depend on indirect mappings and scarce 3D isotropic data for training, primarily focusing on two-dimensional enhancements rather than achieving genuine three-dimensional isotropy. We introduce ``SIMPLE,'' a Simultaneous Multi-Plane Self-Supervised Learning approach for isotropic MRI restoration from anisotropic data. Our method leverages existing anisotropic clinical data acquired in different planes, bypassing the need for simulated downsampling processes. By considering the inherent three-dimensional nature of MRI data, SIMPLE ensures realistic isotropic data generation rather than solely improving through-plane slices. This approach's flexibility allows it to be extended to multiple contrast types and acquisition methods commonly used in clinical settings. Our experiments on two distinct datasets (brain and abdomen) show that SIMPLE outperforms state-of-the-art methods both quantitatively using the Kernel Inception Distance (KID), semi-quantitatively through radiologist evaluations, and qualitatively through Fourier domain analysis. The generated isotropic volume facilitates more accurate volumetric analysis and 3D reconstructions, promising significant improvements in clinical diagnostic capabilities.

Figures (6)

• Figure 1: Our approach integrates ATME as a single-plane deep-learning-based super-resolution model solano2023look and employs a simultaneous multi-plane super-resolution model to achieve isotropic resolution. The red arrows represent components utilized exclusively during the training phase in the model architecture. In contrast, the blue arrow denotes the primary path of the model, used during both the training and inference phases.
• Figure 2: Model Architecture: The dashed red volume denotes the linearly interpolated volume. The blue-, green-, and orange-framed images represent the ATME-generated high-quality coronal, axial, and sagittal slices, respectively. Red, blue, and yellow dashed arrows indicate sampling along the coronal, axial, and sagittal planes, respectively. Gray dashed arrows show the ATME flow and green dashed arrows highlight the connections to consistency loss.
• Figure 3: Multi-plane view of slices sampled from the isotropic MRI volume generated by SIMPLE in comparison to five competitive methods (from left to right): linear interpolation on anisotropic axial volume, linear interpolation on anisotropic coronal volume, ATME on anisotropic axial volume, ATME on anisotropic coronal volume, averaged ATME on coronal and axial planes of anisotropic coronal volume, SMORE4 on anisotropic coronal volume, and SIMPLE on anisotropic coronal volume. SIMPLE produced higher-quality slices simultaneously on all planes compared to the other methods, which produced high-quality slices along the super-resolution plane.
• Figure 4: Straight multi-planar reconstructions (MPR) of the terminal ileum in Crohn's disease patients for three cases based on coronal volumes: (a) anisotropic volume, (b) linearly interpolated isotropic volume, (c) averaged ATMEs isotropic volume, (d) SMORE4 isotropic volume, and (e) SIMPLE isotropic volume. The MPRs generated from the SIMPLE reconstructed volume were less prone to artifacts stemming from the anisotropic MRI acquisition.
• Figure 5: Fourier representation of axial and coronal slices generated from an anisotropic axial and coronal volumes, accordingly. SIMPLE's Fourier representation shows a more consistent distribution of the Fourier coefficients across the two dimensions of the image, whereas other methods images exhibit sampling artifacts. Compared to the closest slice from the axial or coronal GT volume, SIMPLE produces a similar Fourier representation.