PaperlayerView exact revision on arXiv

Source-linked AI summary

Anatomy-Aligned Surface Field Learning for Myocardial Reconstruction from Sparse Short-Axis Cine MRI

Xiaohan Yuan, Xuan Yang, Qingya Li, Yangang Wang, Lei Li

arXiv:2609.29825v1cs.CV

TL;DR

Sparse through-plane sampling in routine SAX cine MRI makes dense, anatomically consistent myocardial reconstruction challenging. This paper completes surface-coordinate fields on a shared anatomy-aligned UV domain, reporting best surface accuracy across three datasets and close agreement in ventricular function.

  • Problem

    Sparse through-plane sampling in routine SAX cine MRI leaves limited geometric evidence for reconstructing smooth, anatomically plausible myocardial surfaces.

    Source: §1 · p. 1

  • Method

    The framework unfolds epicardial and endocardial surfaces onto a shared circumferential-longitudinal UV domain and completes coordinate fields conditioned on encoded sparse SAX contours.

    Source: §1 · p. 2, §3.2.2 · p. 4

  • Results

    The method achieved the best surface accuracy across ACDC, M&Ms, and M&Ms-2, with EDV and LVEF MAEs of 3.3 mL and 1.1%, respectively.

    Source: §5 · p. 10

  • Takeaways & Limitations

    The canonical UV domain also serves as a common representation for functional analysis, pathology localization, and generative modeling.

    Source: §5 · p. 10

    Separate models were trained per dataset, the formulation covers left-ventricular myocardium, and reconstruction quality still depends on the canonical template and UV mapping.

    Source: §5 · p. 10

Abstract

from arXiv · show

Patient-specific 4D myocardial reconstruction from cine MRI supports quantitative functional assessment, regional motion analysis, and simulation-based modeling. However, routinely acquired short-axis (SAX) cine MRI is sparsely sampled along the through-plane direction, making dense and anatomically consistent surface reconstruction challenging. In this study, we propose an anatomy-aligned surface learning framework that parameterizes the epicardial and endocardial surfaces on a shared circumferential-longitudinal UV domain. This formulation converts irregular 3D reconstruction into structured coordinate-field completion with explicit correspondence across subjects and cardiac phases. Sparse SAX contours are encoded as UV observation fields, coverage-aware sampling improves robustness to incomplete slice coverage, and topology- and distortion-aware learning preserves circumferential continuity and local surface quality. Experiments on three public cine MRI datasets showed that the proposed method consistently outperformed representative mesh-based and implicit reconstruction approaches, achieving overall Chamfer distances of $2.887$~mm on ACDC, $2.641$~mm on M\&Ms, and $2.810$~mm on M\&Ms-2. The reconstructed sequences also preserved ventricular function, with end-diastolic volume and ejection fraction errors of $3.3$~mL and $1.1 \%$, respectively. These results demonstrate that anatomy-aligned UV learning provides an accurate, efficient, and correspondence-aware representation for sparse cine MRI reconstruction and myocardial modeling. The source code will be available at https://github.com/yuan-xiaohan/SAX2MyoSurf.

1. Introduction

Sparse short-axis cine MRI provides limited through-plane evidence, making dense, anatomically plausible myocardial reconstruction challenging. The paper organizes both surfaces and sparse observations on a shared circumferential-longitudinal UV domain, then completes them as coordinate fields.

1. Introduction

  • Sparse SAX contours provide limited through-plane evidence, challenging smooth, anatomically plausible myocardial surface reconstruction.Surface representations also support spatial correspondence, regional functional mapping, and simulation-based modeling.Source: §1 · p. 1
  • Prior mesh methods deform templates, while implicit methods predict continuous fields; both leave a gap in organizing sparse observations through intrinsic myocardial coordinates.The paper identifies joint use of a surface domain for sparse SAX observations and dense reconstruction targets as underexplored.Source: §1 · pp. 1–2
  • A shared circumferential-longitudinal UV domain turns irregular 3D reconstruction into structured coordinate-field learning with correspondence across subjects and cardiac phases.The u coordinate follows the circumferential direction, while v follows the longitudinal base-to-apex direction.Source: Figure 1 · p. 1, §1 · p. 2
  • Sparse SAX contours are encoded as surface-aligned UV observation fields, with coverage-aware sampling addressing incomplete slice coverage.Sampling is applied to the physical contours before UV encoding.Source: §1 · p. 2
  • The completion network combines circumferential wrap convolution with region-focused geometric regularization for coherent epicardial and endocardial reconstruction.The framework is validated on three public cine MRI datasets.Source: §1 · p. 2

2. Related work

Prior cardiac reconstruction methods use explicit surfaces, implicit fields, or generative models, while anatomical coordinates and canonical UV representations have largely served other purposes. The paper situates sparse-SAX reconstruction at the intersection of these approaches.

2.1. 3D/4D Cardiac Reconstruction from Cine MRI

  • Explicit surface methods often rely on multi-view imaging, dense supervision, or strong geometric priors, limiting applicability to sparse SAX-only acquisitions.Examples include template and mesh deformation, point-cloud deformation, differentiable slicing, graph refinement, and statistical shape constraints.Source: §2.1 · p. 2
  • Implicit cardiac methods use continuous fields or deformable domains, whereas generative models emphasize distribution learning rather than sparse-observation reconstruction with cross-subject correspondence.The paper contrasts these approaches with directly lifting sparse SAX contours into surface-aligned observation fields for UV completion.Source: §2.1 · p. 2

2.2. Anatomical Coordinates and Canonical Surface Representations

  • Anatomical coordinates support regional analysis and cross-model comparison, but are typically used for post-processing, simulation, or registration rather than primary sparse-observation reconstruction.Canonical UV methods in vision and graphics support dense correspondence and alignment, but largely target natural images or dense surface and texture supervision.Source: §2.2 · p. 3

3. Method

The method represents myocardial surfaces and sparse SAX observations on a shared anatomical UV domain, then completes the fields with topology- and distortion-aware learning. Fixed UV connectivity converts the completed fields into explicit meshes.

3. Method

  • The framework transforms SAX contours into surface-aligned observation fields, completes epicardial and endocardial UV fields, and reconstructs explicit meshes with fixed connectivity.Its three components are shared-domain representation, contour-to-UV encoding with coverage-aware training, and topology- and distortion-aware completion.Source: §3 · p. 3

3.1. Anatomy-Aligned UV Surface Representation

  • A shared circumferential-longitudinal UV domain maps each location to consistent anatomy across subjects and phases, while fixed connectivity converts valid grid locations into triangular meshes.The two surfaces share a centerline, septal reference, and seam convention.Source: §3.1 · p. 3

3.2. Sparse SAX Contour-to-UV Observation Modeling

  • Sparse SAX contours are transformed into observation fields on the same anatomy-aligned domain, with coverage-aware slice sampling intended to improve robustness to varying longitudinal coverage.The method addresses discontinuous contour observations before completing the surface fields.Source: §3.2 · p. 3, Figure 2 · p. 4

3.2.1. Contour-to-UV Observation Encoding

  • For each UV location, the method uses its template point as an anatomical query, retrieves nearby contour candidates, and aggregates slice contributions with distance-dependent soft weights.The selected patient-specific positions are drawn from rigidly aligned contours.Source: §3.2.1 · p. 4

3.2.2. Coverage-Aware Slice Sampling

  • Training samples physical SAX contours under full, missing basal or apical, middle-only, and uniformly sparse coverage patterns before UV encoding.This exposes the network to varying slice coverage and improves robustness to incomplete observations.Source: §3.2.2 · pp. 4–5

3.3. Topology- and Distortion-Aware Surface Completion

  • A shared encoder processes the observation fields and surface-specific decoders reconstruct complete epicardial and endocardial coordinate fields for mesh conversion.The resulting fields are converted to triangular meshes using fixed canonical UV-grid connectivity.Source: §3.3 · p. 5, §3.3.2 · p. 5

3.3.1. Circumferentially Periodic Feature Learning

  • Circumferential wrap convolution uses the UV seam to aggregate features across anatomically adjacent lateral boundaries, maintaining circumferential continuity.The method applies periodic padding along u and conventional padding along v.Source: §3.3.1 · p. 5

3.3.2. Region-Focused Geometric Learning

  • Region-focused geometric losses emphasize a precomputed core of metric compression or UV-grid conflicts and balance its penalties against surrounding regions.The base objective smooths UV coordinate fields, preserves shared epicardial-endocardial vertices, and constrains surface normals and face inversions; region-wise balancing prevents larger regions from dominating the focus core.Source: §3.3.2 · p. 5

4.1. Datasets and Experimental Protocols

The experiments used three public cine MRI datasets with distinct cohort sizes and subject-level splits. Additional data supported scar localization and disease-conditioned motion generation.

4.1. Datasets and Experimental Protocols

  • The reconstruction cohorts comprised 150 ACDC cases, 340 M&Ms cases, and 350 M&Ms-2 cases, with subject-level splits of approximately 70% training and 30% testing.Twenty percent of training subjects were reserved for validation, and test sets remained unseen until final evaluation.Source: §4.1 · p. 5
  • CineMyoPS contributed 64 two-center cases with cine MRI and manual LV, myocardium, and scar annotations for downstream scar localization.The reconstruction model generated UV-domain sequences, and abnormal motion features were used for scar prediction.Source: §4.1 · pp. 5–6
  • Tables 1 and 2 report, respectively, reconstruction comparisons and LV myocardium segmentation comparisons from reconstructed surfaces intersected with SAX planes.Table 1 reports mean (standard deviation); Table 2 identifies best and second-best results.Source: Table 1 · p. 6, Table 2 · p. 6
  • Disease-conditioned motion generation used reconstructed UV sequences from 270 NOR, DCM, and HCM cases, split into 216 training and 54 held-out cases.Source: §4.1 · p. 6

4.2. Reference Mesh Construction and Evaluation

The evaluation uses consistent patient-specific reference meshes and measures surface, SAX-plane, cross-phase motion, and functional agreement.

4.2. Reference Mesh Construction and Evaluation

  • Reference meshes with consistent topology and vertex correspondence were generated offline from multi-phase SAX contours using a cardiac atlas, global alignment, non-rigid refinement, and temporal smoothing.Source: §4.2 · p. 6
  • Surface agreement was measured by symmetric Chamfer distance and F-score at a 2-mm tolerance, while SAX-plane agreement used Dice, ASSD, and HD95.SAX-plane metrics were computed after intersecting reconstructed surfaces with the corresponding segmentations.Source: §4.2 · p. 6
  • Cross-phase motion was evaluated with ED-relative endpoint error at corresponding UV locations, and function with EDV and LVEF.Source: §4.2 · p. 6
  • Ablations additionally used UV MSE over the valid UV domain and Edge Ratio P95 for high-percentile local mesh distortion relative to the reference surface.Source: §4.2 · p. 6

4.3. Implementation

The implementation represents both myocardial surfaces on a canonical UV grid and trains a shared-encoder, surface-specific-decoder network separately for each dataset.

4.3. Implementation

  • Epicardial and endocardial coordinate fields were represented on a 256×256 canonical UV grid and concatenated into a six-channel input.Source: §4.3 · p. 6
  • A seam-aware 2D U-Net used a shared encoder and separate surface-specific decoders, with full- and incomplete-coverage samples generated by coverage-aware sampling.Source: §4.3 · p. 6
  • Separate models were trained from scratch for ACDC, M&Ms, and M&Ms-2 on one RTX 4090 GPU, with checkpoints selected using validation data only.Training lasted 48,000, 75,000, and 80,000 steps for ACDC, M&Ms, and M&Ms-2, respectively.Source: §4.3 · p. 6

4.4. Comparison Study

The method achieved the strongest reported surface reconstruction results across three datasets and also produced accurate, fast SAX-plane evaluations. Qualitative comparisons show recovery across disease-specific morphologies.

4.4. Comparison Study

  • 2.887 mm on ACDC, 2.641 mm on M&Ms, and 2.810 mm on M&Ms-2 were the method’s Chamfer distances, and it significantly outperformed the best baseline on both reported surface metrics across all datasets.Corresponding F@2mm values were 0.796, 0.839, and 0.817, respectively.Source: §4.4 · p. 6
  • The qualitative comparison showed stable recovery of disease-specific ED and ES morphologies across representative NOR, DCM, HCM, and MINF cases.Figure 3 pairs sparse contours and reference surfaces with competing predictions, and shows contours, UV offset maps, and reconstructed surfaces.Source: §4.4 · p. 7, Figure 3 · p. 7
  • On ACDC, the method was comparable to 4DMM for SAX-plane segmentation, while on M&Ms and M&Ms-2 it achieved the best results across all reported metrics.It required 0.8 s per frame, versus more than 30 s for 4DMM and GHDHeart; the shared UV domain represented patient- and phase-specific variation.Source: §4.4 · p. 7
  • Table 3 reports an ACDC ablation study with UV MSE in units of ×10^-3.Source: Table 3 · p. 7

4.5. Ablation Study

The ablation study evaluates alignment, coverage-aware training, region-focused supervision, and WrapConv2D across reconstruction metrics and slice-coverage patterns. Qualitative results show that network completion and targeted components address distinct surface defects.

4.5. Ablation Study

  • Tables 3 and 4 evaluated contributions of centerline alignment, coverage-aware training, region-focused supervision, and WrapConv2D; results were pooled over five coverage patterns and jointly computed for epicardial and endocardial surfaces using UV MSE and CD.The Rigid / Full-only and Centerline / Full-only variants were trained only with full-coverage samples, while the complete model incorporated all proposed components.Source: §4.5 · p. 7
  • Network completion produced complete, smooth surfaces; WrapConv2D improved circumferential seam continuity, while region-focused supervision reduced apical triangle stretching and distortion.Direct UV interpolation produced stripelike artifacts, and converting observed coordinates recovered only locally observed geometry.Source: §4.5 · p. 8

4.6. Utility of the Canonical UV Representation

The canonical UV representation supports functional quantification, scar localization, and generative modeling. Across these applications, reconstructed surfaces retained functional patterns, localized motion abnormalities, and enabled generated disease-related motion sequences.

4.6.1. Myocardial Shape and Functional Quantification

  • EDV MAE was 3.3 mL (R2 = 0.991) and LVEF MAE was 1.1% (R2 = 0.988), with predicted volume and wall-thickening curves following references across the cardiac cycle.The curves preserved disease-specific patterns, including enlarged ventricular volumes and reduced wall thickening in DCM and MINF, and increased thickening in HCM.Source: §4.6.1 · p. 8, Figure 6 · p. 9
  • Coverage-aware training improved all incomplete-coverage settings, with mid-block UV MSE and CD reduced by 91.1% and 56.3%, respectively.It caused only a minor change under full coverage; WrapConv2D and region-focused supervision addressed seam continuity and apical distortion, respectively.Source: §4.6.1 · p. 8, Figure 5 · p. 9

4.6.2. UV-Domain Myocardial Scar Localization from Cine MRI

  • Predicted scar footprints were spatially consistent with references in AHA-17, UV, and reconstructed 3D representations, indicating preserved regional motion characteristics relevant to scar localization.The independent CineMyoPS evaluation derived motion features and mapped abnormal-motion patterns to LGE-derived scar regions.Source: §4.6.2 · pp. 8–9

4.6.3. UV-Domain Generative Modeling

  • Motion latents distinguished NOR, DCM, and HCM distributions, while generated sequences achieved an overall CD of 3.929 mm and preserved characteristic LV cavity-volume patterns.The generated sequences comprised 25 frames for all held-out cases.Source: §4.6.3 · p. 9

5. Discussion and Conclusion

The discussion concludes that shared UV coordinates support accurate, efficient reconstruction and several downstream applications. The study also identifies dataset, anatomical-scope, template, and evaluation limitations.

5. Discussion and Conclusion

  • The framework achieved the best surface accuracy across ACDC, M&Ms, and M&Ms-2, with inference time of 0.8 s per frame and EDV and LVEF MAEs of 3.3 mL and 1.1%.The authors report particularly strong performance on the more heterogeneous M&Ms cohorts.Source: §5 · pp. 9–10
  • Coverage-aware sampling improved reconstruction under missing basal, apical, mid-block, and uniformly sparse observations; seam and apical constraints addressed localized geometric defects.The discussion notes that global surface-distance metrics alone do not characterize quality in parameterization-sensitive regions.Source: §5 · p. 10
  • Shared UV correspondence supported mapping regional motion reductions to LGE-derived scar regions and modeling motion with VAE and latent diffusion architectures.The generative trajectories retained disease-related motion information, surface geometry, and global ventricular function.Source: §5 · p. 10
  • Separate dataset-specific models, LV-only focus, dependence on the canonical template and UV mapping, qualitative scar evaluation, and a small generative cohort bound the current evidence.The generative cohort included only NOR, DCM, and HCM cases.Source: §5 · p. 10
Loading 2609.29825v1…