PaperlayerView exact revision on arXiv

Source-linked AI summary

MRI-based Deep Radiomic Phenotyping of Neuromuscular Disorders: A Topology-driven Characterization

Martyna Żur, Łukasz Piórecki, Marek Socha, Jordi Diaz-Manera, Jose Verdu Diaz, Volker Straub, Rossella Tupler, Joanna Polańska

arXiv:2608.24415v1cs.CV

TL;DR

Muscle MRI assessment needs spatially informative biomarkers because global volumetry can miss disease-specific infiltration architecture and asymmetry. This study develops topology-driven 3D radiomic descriptors and finds substantial discriminative effect sizes for genetic phenotypes.

  • Problem

    Global fat volumetry can miss structural topology, pathological infiltration patterns, and asymmetric genotypic signatures relevant to distinguishing neuromuscular disorders.

  • Method

    The framework segments moderate, severe, and general fat with a GMM and extracts interpretable 3D morphometric, geometric, interface, and graph-topological features.

  • Results

    Substantial effect sizes and post-hoc analyses showed that global 3D descriptors discriminate specific genetic phenotypes through spatial and topological markers.

  • Takeaways & Limitations

    Topology-driven descriptors complement standard clinical volumetry by representing heterogeneous lipodegeneration as interpretable muscle architecture.

  • Takeaways & Limitations

    The method depends on segmentation quality, may be distorted by anisotropic MRI, lacks external validation, and is not strictly adjusted for disease stage.

Abstract

from arXiv · show

Quantitative assessment of muscle MRI is crucial for monitoring neuromuscular disorders (NMD). This study introduces an automated radiomic phenotyping framework based on original features engineered across five main architectural domains: quantitative morphometry, spatial distribution, geometric shape, interactions between progressive fat replacement stages, and graph-based topology. Utilizing 1184 MRI scans from the CoMPaSS-NMD project, we map the complex 3D architecture of heterogeneous intramuscular lipodegeneration into objective, morphologically interpretable biomarkers. We introduce a graph-based skeletonization of fat infiltrates to quantify muscle architectural changes, establishing a multi-dimensional extension of traditional, spatially-agnostic volume metrics by mapping topological networks across the entire 3D muscle volume. Statistical screening via non-parametric Kruskal-Wallis analysis confirmed the discriminative power of these novel descriptors across the genetic hierarchy. Notably, topological network metrics (e.g., SF1_Skel_Nodes, $ε^2$ = 0.2656) and interface dynamics metrics (e.g., SF2_To_SF1_Dist_Min, $ε^2$ = 0.2092) demonstrated substantial effect sizes, providing deeper structural insights than classical volumetric assessments. Post-hoc pairwise evaluations and UMAP projections further indicated the capability of these topological and 3D geometric invariants to capture disease-specific macroscopic infiltration patterns. These results demonstrate that global architectural features represent a highly promising class of biomarkers for differential diagnosis, offering new avenues for tracking longitudinal disease dynamics in neuromuscular diagnostics. The developed automated feature extraction pipeline is integrated and available within the MUSCAT (MUSCle fAt Topology) library.

1 Introduction

MRI is central to evaluating progressive muscle degeneration, but overlapping NMD symptoms make precise morphological characterization important for treatment decisions. Traditional visual grading and volumetric metrics motivate more continuous, spatially informative assessment.

  • 1 Introduction: MRI evaluates progressive muscle degeneration by distinguishing functional tissue from moderate and severe fat replacement.These replacement stages correspond to intermediate and high radiological intensities, respectively.
  • 1 Introduction: Overlapping symptoms across neuromuscular disorders make precise replacement morphology important for differential therapeutic decisions.The passage emphasizes that similar clinical presentations may require fundamentally different interventions.
  • 1 Introduction: The Mercuri scale standardized radiological reporting but provides semi-quantitative visual assessment.It remains a cornerstone of clinical diagnostics.
  • 1 Introduction: Fat Fraction mapping provides reproducible volumetric measurements but compresses complex 3D muscle architecture into one scalar.Distinct spatial arrangements can therefore share the same global fat fraction.

2 State of the Art: Feature Extraction in Muscle MRI

Muscle MRI assessment has progressed from categorical visual grading toward quantitative volumetry, but both approaches leave important structural information unresolved. The key limitation is that global fat burden does not encode infiltration topology or marked asymmetry.

  • 2.1 Visual Grading and Quantitative Volumetry: The Mercuri scale lacks sensitivity for sub-clinical disease monitoring because of its discrete, categorical design.It remains important for standardized radiological reporting, but cannot provide continuous measurements.
  • 2.1 Visual Grading and Quantitative Volumetry: Fat Fraction mapping reduces inter-observer bias but remains limited to global volumetry.It cannot distinguish focal fat accumulation from structurally invasive networks.
  • 2.1 Visual Grading and Quantitative Volumetry: Averaging left and right limb volumetric scores can mask the extreme patchy asymmetry associated with DUX4 mutations.This motivates advanced geometric profiling that preserves spatial differences.

2.2 Radiomics and the "Bag-of-Pixels" Problem

Classical radiomics and manual 3D annotation are poorly suited to representing heterogeneous muscle pathology across complete volumes. Their spatial blindness and annotation burden encourage reduced, single-slice analyses.

  • 2.2 Radiomics and the "Bag-of-Pixels" Problem: GLCM- and GLSZM-based radiomics aggregate pixels into global matrices, making them spatially blind.They cannot distinguish scattered fat islands from a unified invasive network.
  • 2.2 Radiomics and the "Bag-of-Pixels" Problem: Manual 3D annotation is operator-dependent and impractically time-consuming for large MRI datasets.This burden contributes to restricting many studies to a single mid-thigh cross-section.
  • 2.2 Radiomics and the "Bag-of-Pixels" Problem: Atlas-based and active-contour segmentation methods often fail in advanced neuromuscular disorders.Severe fatty infiltration and atrophy obscure fascia and inter-muscular boundaries, undermining healthy anatomical atlases.

2.4 Multi-Center Heterogeneity and Scale Information Leakage

Multi-center clinical annotations can introduce scale-dependent artifacts into diagnostic modeling, while data-hungry CNNs face annotation scarcity and scanner generalizability challenges. These limitations motivate continuous, image-driven features with more explicit structural representation.

  • 2.4 Multi-Center Heterogeneity and Scale Information Leakage: Different institutions may use incompatible visual grading schemes, such as 4-point versus 6-point scales.This heterogeneity complicates aggregation of retrospective clinical annotations.
  • 2.4 Multi-Center Heterogeneity and Scale Information Leakage: Scale information leakage occurs when models learn institution- or disease-associated grading artifacts instead of underlying pathophysiology.The problem arises when rare diseases are disproportionately annotated using a distinct historical scale.
  • 2.4 Multi-Center Heterogeneity and Scale Information Leakage: Continuous image-driven spatial metrics can bypass human quantization bias in multi-center data.This is presented as a critical need for automated diagnostic pipelines.
  • 2.4 Multi-Center Heterogeneity and Scale Information Leakage: CNNs face a data-hunger barrier in rare diseases and may generalize poorly across scanners with different hardware or field strengths.Models can overfit to hardware-specific signal intensities and contrast variations.

2.6 The Tractography Limitation in Lipodegeneration

DTI and tractography become unreliable in advanced muscular dystrophies as fat replacement reduces water signal and introduces partial-volume and SNR problems. The framework instead models the 3D architecture of invasive fat networks using morphologically grounded graphs.

  • Advanced lipodegeneration undermines DTI-based architectural tracking through diminished water signal, partial-volume effects, and reduced SNR.
  • Graph-Radiomics represents fat infiltrates as physical skeletal infrastructure rather than abstract statistical clusters.
  • This morphological graph representation connects structural complexity directly to macroscopic tissue anatomy and clinical interpretability.

2.8 Statistical Rigor: Beyond P-Value

The study argues that radiomic discovery should assess clinical magnitude rather than rely solely on statistical significance. It therefore uses Epsilon-squared effect sizes to distinguish meaningful biomarkers from statistical noise.

  • Epsilon-squared effect size is used to evaluate the clinical magnitude of engineered radiomic findings beyond p<0.05 screening.
  • The approach addresses the sensitivity of p-values to sample size and their inability to represent clinical magnitude.
  • Effect-size screening is intended to ensure proposed features are not merely statistical noise.

3 Methods: Feature Engineering and Mathematical Basis

The methods combine automated segmentation, multidomain 3D feature engineering, and graph-based skeletonization to characterize muscle fat architecture. Features quantify burden, spatial organization, shape, interface proximity, fragmentation, and network topology across standardized fat masks.

  • Study Population and Tissue Segmentation: 1184 MRI scans from five genetic phenotypes were processed with automated intensity-based segmentation in MUSCAT.The cohort included DYSF, CAPN3, GNE, DMPK, and DUX4 phenotypes.
  • Study Population and Tissue Segmentation: A GMM separated functional muscle from pathology into SF1, SF2, and combined GF masks, with bilateral leg features averaged per patient.SF1 denotes moderate replacement, SF2 severe replacement, and GF their union.
  • Feature Engineering: The framework spans five architectural domains and uses Mutual Information to retain 29 unique, orthogonal spatial biomarkers.The formulations are applied independently to GF, SF1, and SF2.
  • Volumetric Macro-structure: Relative disease burden normalizes voxel-count volumes by physical spacing and anatomical compartment volume to reduce anthropometric bias.Absolute volume is not used as a standalone biomarker.
  • Morphological Fragmentation: Three-dimensional 26-connectivity identifies discrete fat islands while limiting artificial fragmentation of thin diagonal infiltrates.
  • Geometric Shape and Interface Dynamics: PCA eigenvalues quantify macroscopic shape, while interface distances measure proximity between severe and moderate fat masks.The shape analysis uses ordered eigenvalues; interface distance uses Euclidean boundary distances.
  • 3D Graph-based Topology: Morphological skeletonization reduces fat masks to one-voxel-thick medial lines while preserving topology, then maps them into graphs of vertices and edges.Network length uses edge lengths and MRI spacing; cycles use edges, vertices, and disconnected components.

4 Results

The engineered biomarkers were screened for global variance and pairwise phenotype differentiation, then evaluated through UMAP and feature-gradient projections. Topological and geometric descriptors captured disease-specific 3D architectural patterns, with branching structure supporting separation of specific cohorts.

  • Kruskal-Wallis testing evaluated whether at least one disease group differed from the others across engineered biomarkers.
  • The feature dictionary organizes engineered 3D spatial biomarkers by architectural domain, extracting most features independently for GF, SF1, and SF2 masks.
  • ε2 = 0.2656 for SF1_Skel_Nodes and ε2 = 0.2542 for GF_Skel_Nodes, indicating substantial differentiation across genetic phenotypes.
  • Dunn’s post-hoc tests showed that topological and geometric features isolated specific phenotypes across the entire 3D muscle volume.
  • r > 0.88 for GF_Skel_Nodes and SF1_Skel_Nodes in comparisons isolating DUX4 from DMPK and DYSF.
  • UMAP used 28 spatial biomarkers with ε2 ≥0.01, while gradient maps linked manifold separation to orientation, skeletal branching, and fragmentation signatures.
  • The isolated DUX4 cluster exhibited extreme density of topological branching nodes, connecting its manifold separation to complex multidirectional 3D network structure.

5 Discussion

The framework extends muscle MRI assessment from volumetric fat estimation to interpretable 3D architectural profiling, capturing topology, geometry, fragmentation, and spatial dynamics across disease phenotypes. These continuous biomarkers separate difficult-to-distinguish phenotypes and may support longitudinal tracking, although segmentation, acquisition anisotropy, cohort heterogeneity, and absent external validation constrain interpretation.

  • Clinical phenotype translation: Global fragmentation metrics, including SF2_BlobSize_Median, isolate phenotypes characterized by patchy, multifocal infiltrates associated with early muscle edema.The paper relates this pattern to edema preceding irreversible fat replacement.
  • Clinical phenotype translation: Topological markers such as GF_Skel_Nodes distinguish cohesive tissue replacement from highly branched degeneration and isolate DUX4 patients in a branching-rich cluster.The DUX4 cluster is contrasted with continuous LGMD patterns.
  • Clinical phenotype translation: Interface dynamics, center-of-mass, blob-size, and geometric features separate genetically related phenotypes that standard volumetry and visual grading often confound.SF2_To_SF1_Dist_Min and SF1_CoM_Z capture tissue proximity and spatial positioning, while 3D geometric signatures differentiate allied variants.
  • Dimensionality and morphological continuum: UMAP based on discriminative global features groups cohorts by spatial-biomarker similarity, isolating DUX4 and GNE patterns while retaining a continuous main cluster for overlapping LGMD phenotypes.The framework interprets this organization as reflecting both phenotypic divergence and macroscopic similarity.
  • Dimensionality and morphological continuum: Five topological groups expose distinct architectural signatures, including horizontal propagation, pooled macroscopic lakes, and elongated streak-like infiltrates.These patterns are represented by GF_MainAxis_Z, SF1_BlobSize_Max, and GF_Elongation, respectively.
  • Longitudinal tracking: Topological variables such as GF_Skel_Cycles and SF2_To_SF1_Dist_Min offer candidate measures of spatial disease dynamics beyond static Fat Fraction.The paper frames these metrics as potentially useful for longitudinal tracking.
  • Algorithmic robustness and interpretability: The method is intrinsically interpretable because spatial vectoring outputs physical coordinates and graph measures rather than relying on post-hoc explanations.The authors contrast direct center-of-mass coordinates with approximate black-box explanations.
  • Limitations: Interpretation is limited by segmentation quality, anisotropic MRI resolution, missing external validation, unadjusted disease-stage heterogeneity, and computational cost.Segmentation errors propagate to skeletal graphs, while acquisition anisotropy can distort node and cycle counts.

6 Conclusion

The study proposes topology-driven MRI biomarkers that capture spatial heterogeneity and global 3D architecture beyond conventional volumetry. Statistical and post-hoc analyses support their use for distinguishing genetic phenotypes, while clinical deployment remains subject to data and computational constraints.

  • Topology-driven descriptors model lipodegeneration through structural network complexity, fragmentation, and interface dynamics, complementing standard Fat Fraction volumetry.The approach addresses the spatial compression of global Fat Fraction and the limitations of classical bag-of-pixels radiomics.
  • Global 3D descriptors yielded substantial effect sizes in statistical evaluation.
  • Post-hoc analyses found spatial markers capable of isolating genetic phenotypes by their unique macroscopic infiltration patterns.
  • The morphological approach provides a transparent, geometrically accurate foundation for deep radiomic phenotyping and differential diagnosis.
  • Clinical deployment requires addressing incomplete disease-stage data and the greater computational cost of 3D distance and graph-network calculations.Exact disease-stage adjustment was limited by unavailable progression data, while the topology modules are more computationally demanding than standard volumetric assessments.

Code Availability

The automated feature extraction pipeline, including its 3D graph-based topology modules, is integrated into the MUSCAT library. The code and feature-engineering algorithms are available from the corresponding author upon reasonable request.

  • The automated feature extraction pipeline and 3D graph-based topology modules were developed as part of the MUSCAT library.
  • The codebase and associated feature-engineering algorithms are available from the corresponding author upon reasonable request.
Loading 2608.24415v1…