AtlasGS: Brain MRI Spatial Resolution Harmonization With Shared Gaussian Geometry

AtlasGS introduces a novel two-stage framework that leverages shared Gaussian geometry learned from isotropic structural scans to achieve state-of-the-art spatial resolution harmonization and arbitrary-view generation for sparse multi-contrast brain MRI modalities, demonstrating superior reconstruction fidelity across diverse datasets and pathological conditions.

The Gist

The Problem: The "Mismatched Puzzle"

Imagine you are trying to build a 3D model of a human brain using MRI scans. In a typical hospital, doctors take several different types of pictures (scans) of the same brain to check for different things:

• Scan A (T1): A high-quality, 3D picture where every slice is thin and clear, like a stack of crisp, thin paper.
• Scan B, C, D (T2, FLAIR, etc.): These are taken quickly to catch specific diseases, but they are "chunky." They are like a stack of thick cardboard slices. Because the slices are thick, you lose detail between them, and the image looks blurry or blocky when you try to look at it from the side.

The problem is that these "chunky" scans don't line up perfectly with the "crisp" scan. If you try to combine them or zoom in on the chunky ones using standard computer tricks (like stretching a low-res photo), the image gets blurry, or the computer invents fake details (hallucinations) that aren't really there. This makes it hard for doctors to compare the different scans or see the brain clearly from any angle.

The Solution: AtlasGS (The "Shared Skeleton")

The authors created a new tool called AtlasGS. Instead of trying to fix each blurry scan individually, they built a shared "skeleton" for the brain that all the scans can use.

Think of it like this:

1. The Skeleton (Stage 1): First, the computer looks at the high-quality "crisp" scan (Scan A). It builds a 3D framework made of millions of tiny, glowing dots (called Gaussians). These dots map out the exact shape of the brain's anatomy—the curves, the folds, and the boundaries. This is the "Atlas."
2. The Skin (Stage 2): Next, the computer takes the "chunky" blurry scans (Scan B, C, D). It doesn't try to rebuild the shape; the shape is already locked in by the skeleton. Instead, it just paints the "skin" (the color and texture) onto that pre-built skeleton to match what the blurry scan sees.

Because the shape is already perfect (from the crisp scan), the computer can fill in the missing gaps in the chunky scans without guessing or inventing fake brain parts.

How It Works: The "Video Projector" Analogy

The paper uses a technique called Gaussian Splatting. Imagine you have a video projector that can shoot light in 3D space.

• Standard methods try to guess what the missing light should look like between the thick slices, often resulting in a blurry mess.
• AtlasGS sets up a rigid, invisible wireframe (the skeleton) based on the perfect scan. Then, it projects the "chunky" images onto this wireframe. Because the wireframe is solid, the projection stays sharp, even if you zoom in or look at the brain from a weird angle that the original machine never actually took a picture of.

What They Found (The Results)

The researchers tested this on three different groups of people: healthy older adults, healthy young people, and patients with brain tumors (glioblastoma).

• Sharper Images: When they tried to turn the "chunky" scans into high-resolution images, AtlasGS produced much clearer results than standard methods. It was especially good at fixing scans where the slices were very thick (up to 7 times thicker than normal).
• No Fake Tumors: A major issue with old methods is that when they try to sharpen a blurry image of a tumor, they often create fake edges or distort the shape. AtlasGS kept the tumor boundaries sharp and accurate because it was anchored to the real anatomy from the first scan.
• Arbitrary Views: Because the brain is now a 3D object made of these tiny dots, doctors can "slice" the brain in any direction they want (even angles the machine never scanned) and still see a clear, high-quality image.

The Bottom Line

AtlasGS is a way to take a high-quality 3D map of a brain and use it to fix blurry, low-quality 2D slices from other scans. It acts like a sturdy scaffold that holds the brain's shape steady, allowing the computer to paint in the missing details accurately. This helps doctors see the brain clearly from any angle without inventing fake structures, making it easier to compare different types of MRI scans.

Technical Summary

Technical Summary: AtlasGS – Brain MRI Spatial Resolution Harmonization With Shared Gaussian Geometry

Problem Statement
Clinical magnetic resonance imaging (MRI) protocols are inherently spatially heterogeneous. While neuroimaging protocols typically include a high-resolution isotropic structural scan (e.g., T1-weighted), disease-targeted sequences (such as T2-weighted, FLAIR, DWI, and ASL) are often acquired with anisotropic slice thickness, varying resolutions, and non-standard trajectories. This variability creates inconsistent geometric representations across contrasts, hindering reliable multimodal joint analysis, cross-contrast visualization, and arbitrary-view reconstruction. This is particularly challenging in retrospective datasets where acquisition parameters cannot be easily simulated or standardized. Conventional approaches, such as interpolation, produce over-smoothed images and fail to align geometric representations across contrasts, while deep learning-based super-resolution models often struggle with the wide variability and individualized requirements of clinical datasets.

Methodology: AtlasGS Framework
The authors propose AtlasGS, a Gaussian Splatting (GS)-based framework that utilizes a shared Gaussian geometry to harmonize spatial resolution across multi-contrast MRI. The core hypothesis is that an explicit, subject-specific anatomical scaffold learned from the isotropic structural scan can be reused to fit the appearance of target modalities acquired with sparse slices.

The framework operates via a two-stage training strategy:

1. Stage I: Structural Learning (Geometric Scaffold)

   • An explicit Gaussian scaffold is learned from the high-resolution isotropic T1-weighted scan ($I_{T1}$).
   • The model employs a 2D+t Gaussian field where each primitive $G_i$ includes geometric parameters (spatial center, covariance, through-plane variance) and deformation functions to model nonlinear anatomical continuity along the through-plane axis.
   • Topology-aware Gaussian densification is applied using Persistence Homology (PH) constraints to preserve anatomical integrity and suppress structural artifacts.

2. Stage II: Modality Fitting (Appearance Adaptation)

   • The learned geometric parameters are frozen.
   • The model fits modality-specific appearance parameters (opacity and color) to the target low-resolution (LR) modality ($I_{LR}$) using a latent appearance regularization strategy. Appearance is factorized through a low-dimensional latent code to ensure stable cross-modality fitting.
   • Slab Integration: To simulate the physics of thick-slice acquisition during training, the continuous Gaussian field is integrated along the through-plane coordinate to generate predicted LR slices, which are compared against the ground truth.

3. Low-Resolution (LR) Consistent Fusion

   • To balance the structural prior from the T1 scaffold with the signal fidelity of the target modality, a voxel-wise LR confidence map is computed.
   • This map fuses the T1-guided reconstruction with the single-modality reconstruction, suppressing high-frequency Gaussian blobs and artifacts while preserving absolute signal values essential for quantitative MRI.

Key Contributions

• Explicit Geometry-Guided Representation: Unlike Implicit Neural Representations (INR) or standard GS methods that do not inherently encode semantic organ geometry, AtlasGS explicitly learns an anatomical prior from the isotropic scan to guide the reconstruction of anisotropic modalities.
• Zero-Shot Through-Plane Super-Resolution: The framework enables arbitrary continuous re-slicing and super-resolution (factors of $\times3, \times5, \times7$) without requiring external training data or paired high-resolution target images.
• Clinical Feasibility: The method addresses the specific challenge of clinical datasets where slice thickness often exceeds 5 mm, demonstrating the ability to mitigate tumor hallucinations common in interpolation-based methods.

Experimental Results
Experiments were conducted on three datasets: UK Biobank (healthy older cohorts), UPenn-GBM (glioblastoma patients), and ABCD (healthy younger cohorts). The evaluation covered multiple modalities (T2w, FLAIR, DWI, ASL) and degradation factors.

• Quantitative Performance: AtlasGS achieved state-of-the-art results in MAE, PSNR, and SSIM across all degradation factors on the UK Biobank dataset. The performance margin widened under severe downsampling (e.g., $\times7$), indicating robustness under sparse supervision.
• Pathology Preservation: In the GBM cohort, the model maintained stable lesion structures (measured by nn-UNet Dice scores), whereas modality-agnostic baselines often distorted pathology or underperformed linear interpolation.
• Generalization: The method successfully generalized to rare modalities with non-integral scaling ratios in the ABCD dataset (DWI and ASL).
• Qualitative Improvements: Visualizations confirmed the generation of continuous high-resolution representations in axial, coronal, and sagittal views, with clear tumor boundaries and the elimination of interpolation-induced hallucinations.

Significance and Claims
The paper claims that AtlasGS establishes explicit geometry-guided representations as a novel, flexible, and interpretable pathway for retrospective multi-contrast MRI harmonization. By leveraging the commonly acquired isotropic T1 scan as a geometric anchor, the framework facilitates:

1. Reliable Clinical Reference Construction: Creating consistent anatomical representations across heterogeneous protocols.
2. Arbitrary-View Generation: Enabling the visualization of target modalities from any angle with strong structural consistency.
3. Potential for In-Plane Super-Resolution: The learned Gaussian scaffold shows potential for enhancing in-plane resolution as well.

The authors note that while the framework is robust under anisotropic sampling, it has not yet been evaluated under motion corruption or significant anatomical deformation, and population-level atlas modeling remains a future direction. The work positions explicit geometry as a lightweight yet powerful anchor for medical image reconstruction, offering a scalable foundation for cross-modal registration, longitudinal analysis, and image-guided intervention.