Anatomy-Aligned Surface Field Learning for Myocardial Reconstruction from Sparse Short-Axis Cine MRI
Authors: Xiaohan Yuan, Xuan Yang, Qingya Li, Yangang Wang, Lei Li
Organizations: Department of Biomedical Engineering, National University of Singapore, Singapore · School of Automation, Southeast University, Nanjing, China
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.
Figures & tables
Figure 1: Overview of the anatomy-aligned UV representation for sparse short-axis (SAX) myocardial reconstruction. The epicardial (epi) and endocardial (endo) surfaces are unfolded onto shared circumferential-longitudinal UV domains, where each location stores the corresponding 3D surface coordinates.
Figure 2: Overall pipeline of the proposed anatomy-aligned UV myocardial surface reconstruction framework. Sparse SAX contours are aligned to the canonical template and encoded as epicardial and endocardial UV observation fields, with coverage-aware sampling used during training. A topology- and distortion-aware UV network completes the patient-specific surface fields through circumferential WrapConv2D and region-focused geometric supervision, followed by explicit mesh reconstruction using fixed UV connectivity.
Method
ACDC
M&Ms
M&Ms-2
Inference
CD (mm) ↓
F@2mm↑
CD (mm) ↓
F@2mm↑
CD (mm) ↓
F@2mm↑
Time (s/frame) ↓
Voxel2Mesh
8.629 (2.773)
0.230 (0.126)
8.325 (2.742)
0.246 (0.120)
8.560 (2.761)
0.246 (0.125)
1.4
PN-GCN
3.183 (0.395)
0.741 (0.092)
2.873 (0.446)
0.808 (0.092)
3.144 (0.838)
0.770 (0.106)
1.0
4DMM
2.983 (0.398)
0.782 (0.067)
2.776 (0.339)
0.820 (0.064)
2.914 (0.387)
0.797 (0.066)
>30
GHDHeart
4.700 (1.060)
0.463 (0.112)
4.528 (1.169)
0.511 (0.131)
4.684 (1.012)
0.479 (0.117)
>30
Ours
2.887 (0.388)
0.796 (0.083)
2.641 (0.381)
0.839 (0.076)
2.810 (0.524)
0.817 (0.076)
0.8
Table 1: Quantitative comparison of reconstruction on ACDC, M&Ms, and M&Ms-2. Results are reported as mean (standard deviation).
Method
ACDC
M & Ms
M & Ms-2
Dice ↑
ASSD (mm) ↓
HD95 (mm) ↓
Dice ↑
ASSD (mm) ↓
HD95 (mm) ↓
Dice ↑
ASSD (mm) ↓
HD95 (mm) ↓
Voxel2Mesh
0.506 (0.200)
2.237 (1.513)
9.171 (5.059)
0.531 (0.162)
2.082 (1.345)
9.615 (4.185)
0.481 (0.132)
3.468 (1.245)
17.586 (4.575)
PN-GCN
0.899 (0.044)
0.329 (0.295)
2.616 (3.368)
0.923 (0.043)
0.280 (0.394)
2.664 (3.733)
0.893 (0.070)
0.398 (0.668)
3.409 (5.186)
4DMM
0.956 (0.024)
0.172 (0.205)
2.386 (3.011)
0.953 (0.028)
0.227 (0.316)
2.765 (3.574)
0.944 (0.031)
0.250 (0.323)
3.203 (4.014)
GHDHeart
0.806 (0.071)
0.790 (0.489)
4.448 (2.898)
0.786 (0.062)
1.216 (0.720)
8.979 (5.262)
0.750 (0.069)
1.256 (0.684)
9.597 (5.397)
Ours
0.939 (0.031)
0.212 (0.213)
2.429 (2.654)
0.965 (0.021)
0.114 (0.204)
1.181 (2.327)
0.949 (0.045)
0.193 (0.512)
1.932 (4.029)
Table 2: Quantitative comparison of left ventricular (LV) myocardium segmentation obtained by intersecting the reconstructed surfaces with the original SAX planes. Best and second-best results are shown in bold and underlined, respectively.
Figure 3: Qualitative comparison of disease-specific myocardial reconstruction: (a) ED and ES surfaces for representative NOR, DCM, HCM, and MINF cases, showing sparse SAX contours, reference surfaces, and predictions from the competing methods. (b) Corresponding SAX contours, UV surface-offset maps, and reconstructed 3D surfaces. For visualization, UV fields show normalized template-relative 3D offsets in RGB. NOR: normal subjects; DCM/HCM: dilated/hypertrophic cardiomyopathy; MINF: myocardial infarction with altered LV ejection fraction.
Figure 4: Qualitative comparison of LV myocardial segmentation obtained by intersecting the reconstructed surfaces with the original SAX planes.
Variant
UV MSE ↓
CD (mm) ↓
Edge Ratio P95 ↓
Rigid / Full-only
1.565 (3.413)
3.569 (1.829)
1.447 (0.187)
Centerline / Full-only
1.491 (3.106)
3.538 (1.718)
1.449 (0.189)
w/o Region Focus
0.499 (0.541)
3.021 (0.571)
1.408 (0.053)
w/o WrapConv2D
0.535 (0.573)
2.999 (0.532)
1.406 (0.052)
Ours
0.487 (0.531)
2.997 (0.546)
1.393 (0.049)
Table 3: Ablation study of the proposed framework on ACDC. UV MSE is reported in units of (×10−3) .
UV MSE ↓
CD (mm) ↓
Pattern
Full-only
Coverage-aware
Full-only
Coverage-aware
Full
0.383
0.408
2.852
2.888
Base-missing
1.380
0.544
3.913
3.089
Apex-missing
0.998
0.520
3.303
2.989
Mid-block
10.868
0.963
8.464
3.702
Uniform-sparse
2.122
0.598
4.302
3.139
Table 4: Reconstruction accuracy under five slice-coverage patterns for models trained with full coverage only and with coverage-aware sampling.
Figure 5: (a) Full-only and coverage-aware reconstructions under five slice-coverage patterns. (b) Representative results for locally irregular contour observations. (c) Comparison of non-learning UV reconstruction and network completion (top), seam continuity without and with WrapConv2D (middle), and apical geometry without and with region-focused supervision (bottom).
Figure 6: Quantitative analysis of myocardial shape and function. (a) Predicted and reference LV volume and global wall-thickening trajectories for the five ACDC diagnostic groups; shaded regions indicate group-wise standard deviations. (b) Agreement between predicted and reference EDV and LVEF across ACDC, M&Ms, and M&Ms-2. ARV: abnormal right ventricle.
Figure 7: Qualitative myocardial scar localization in three representative CineMyoPS cases.
Figure 8: Conditional myocardial motion generation and analysis in the UV domain. (a) Projection of the VAE motion latents onto the first two linear discriminant axes, showing disease-related distributions of NOR, DCM, and HCM. (b) Reference and generated normalized LV cavity-volume curves, where the horizontal axis denotes the normalized cardiac phase and the vertical axis denotes V(t)/V(0) . IQR: interquartile range.
Department of Biomedical Engineering, National University of Singapore, Singapore · School of Automation, Southeast University, Nanjing, China · Department of Medicine, National University of Singapore, Singapore +1
Department of Biomedical Engineering, National University of Singapore, Singapore · School of Automation, Southeast University, Nanjing, China · Department of Medicine, National University of Singapore, Singapore +1