SPARC: Slice-to-volume Pipeline for Automated Reconstruction of gated 3D+time fetal Cardiac MRI

Arnaud Boutillon; Naomi Clarke; Tomas Woodgate; Daniel West; Alina Schneider; Rachael Franklin; Anthony Price; Jo Hajnal

cs.CVAug 19, 2026

SPARC: Slice-to-volume Pipeline for Automated Reconstruction of gated 3D+time fetal Cardiac MRI

Authors: Arnaud Boutillon, Naomi Clarke, Tomas Woodgate, Daniel West, Alina Schneider, Rachael Franklin, Anthony Price, Jo Hajnal, +3 more

Organizations: Biomedical Computing Research Department, School of Biomedical Engineering and Imaging Sciences, King’s College London, London, United Kingdom · Imaging Physics and Engineering Research Department, School of Biomedical Engineering and Imaging Sciences, King’s College London, London, United Kingdom · Early Life Imaging Research Department, School of Biomedical Engineering and Imaging Sciences, King’s College London, London, United Kingdom · Medical Physics and Clinical Engineering, Guys and St Thomas’ NHS Foundation Trust, London, United Kingdom

Abstract

Fetal cardiac MRI (fCMR) provides valuable diagnostic information complementary to echocardiography, particularly for complex congenital heart disease (CHD). Dynamic cine imaging captures cardiac motion essential for assessment of cardiac function; however, the reconstruction of 3D+time cine volumes from 2D+time acquired slices remains challenging due to unpredictable fetal motion and the absence of automated and robust processing tools suitable for clinical deployment. We present the SPARC pipeline (Slice-to-volume Pipeline for Automated Reconstruction of gated 3D+time fetal Cardiac MRI) which combines physics-informed slice-to-volume reconstruction (SVR) of Doppler ultrasound (DUS) gated stacks of slices, assisted by deep learning (DL) models for thoracic segmentation and anatomical reorientation. The proposed SVR algorithm achieves a tenfold reduction in reconstruction time relative to existing frame-wise approaches ( $4.8 \pm 1.0$ vs $49.0 \pm 14.1$ min, $p < 0.0001$ ) while improving the reconstruction quality. Thoracic segmentation performance using ensemble aggregation exceeded inter-rater agreement (Dice $84.7 \pm 3.9\%$ vs $81.4 \pm 7.7\%$ , $p<0.05$ ), while anatomical reorientation achieved a success rate of $90.1\%$ . End-to-end evaluation on a large held-out clinical cohort ( $n = 121$ ) demonstrated fully automatic processing in $82.6\%$ of cases with a mean runtime of $7.1 \pm 1.3$ min, compatible with clinical deployment. The complete SPARC pipeline is publicly available as a Docker container https://hub.docker.com/r/aboutill/sparc and is currently deployed at our institution as a clinical research tool.

Figures & tables

Panel of Figure 3 — Figure 3: Overview of the proposed SPARC pipeline.

Domain	Evaluation	Subjects $n$
Domain	Evaluation	5-fold cross- validation set	Held-out test set	Unique subjects
Source	Thoracic segmentation	116 [483]	23 (92)	141
	Cardiac segmentation	98	22
	Reorientation	91	21
Target	Thoracic segmentation	25 [222]	5 (36)	60
	Reconstruction	–	10

Table 1: Dataset splits for each evaluated task in the source (free-running;

n=141

subjects) and target (Doppler-gated;

n=199

subjects) domains. Numbers of stacks are shown in brackets.

Domain		Ensemble	DSC [%] $\uparrow$	$\textrm{HD}_{95\%}$ [mm] $\downarrow$	ASSD [mm] $\downarrow$
Test	Training	Ensemble	DSC [%] $\uparrow$	$\textrm{HD}_{95\%}$ [mm] $\downarrow$	ASSD [mm] $\downarrow$
Source	Source	✓	$84.9\pm 7.0$	$4.1\pm 3.0$	$1.1\pm 0.7$
Source	Joint	✓	$\mathbf{85.7\pm 6.0}$	$4.1\pm 3.2$	$1.1\pm 0.7$
Target	Source	✓	$79.2\pm 7.3$	$10.8\pm 10.6$	$2.1\pm 2.1$
	Target	✓	$81.8\pm 4.5$	$6.5\pm 2.8$	$1.4\pm 0.4$
	Joint	✓	$84.3\pm 3.9$	$5.9\pm 2.8$	$1.2\pm 0.4$

Table 2: Quantitative evaluation of thoracic segmentation models on held-out test sets from the source (free-running;

m=92

stacks across

n=23

subjects) and target (Doppler-gated;

m=36

stacks across

n=5

subjects) domains shown as mean

\pm

standard deviation. The best performance within each domain is shown in bold (linear mixed-effect model with Bonferroni correction,

p<0.05

Method		van Amerom	Proposed
Model assumptions	Motion states	Frame-wise	Slice-wise
Model assumptions	tPSF	✓	✗
Step complexity	Slice/frame-to-volume registration	$\mathcal{O}{(KT)}$	$\mathcal{O}{(K)}$
Step complexity	Super-resolution reconstruction	$\mathcal{O}{(NT^{2})}$	$\mathcal{O}{(NT)}$
Leave-stack -out	NCC $\uparrow$	$0.89\pm 0.04$	$\mathbf{0.91\pm 0.03}$
Leave-stack -out	NRMSE $\downarrow$	$0.13\pm 0.03$	$\mathbf{0.11\pm 0.02}$

Table 3: Quantitative evaluation of the reconstruction algorithms on the target domain test set (Doppler-gated;

n=10

subjects) shown as mean

\pm

standard deviation. The best performance within each domain is shown in bold (paired

t

-test,

p<0.05

)

Domain		Ensemble	DSC [%] $\uparrow$	$\textrm{HD}_{95\%}$ [mm] $\downarrow$	ASSD [mm] $\downarrow$	CD [mm] $\downarrow$
Test	Training	Ensemble	DSC [%] $\uparrow$	$\textrm{HD}_{95\%}$ [mm] $\downarrow$	ASSD [mm] $\downarrow$	CD [mm] $\downarrow$
Source	Source	✓	$83.4\pm 4.9$	$4.8\pm 1.9$	$1.7\pm 0.4$	$4.2\pm 2.3$
Source	Joint	✓	$83.8\pm 5.0$	$4.7\pm 1.5$	$1.6\pm 0.4$	$4.0\pm 2.3$
Target	Source	✓	$20.0\pm 30.8$	$27.8\pm 12.0$	$14.4\pm 7.7$	$39.5\pm 22.3$
	Target	✓	$73.8\pm 9.0$	$8.2\pm 3.6$	$3.2\pm 1.4$	$12.3\pm 5.2$
	Joint	✓	$71.0\pm 7.1$	$9.2\pm 1.4$	$3.5\pm 0.9$	$11.1\pm 7.6$

Table 4: Quantitative evaluation of cardiac segmentation models on held-out test sets from the source (free-running;

n=22

subjects) and target (Doppler-gated;

n=5

subjects) domains shown as mean

\pm

standard deviation.

Domain		Ensemble	GD [ $\degree$ ] $\downarrow$	CD [mm] $\downarrow$
Test	Training	Ensemble	GD [ $\degree$ ] $\downarrow$	CD [mm] $\downarrow$
Source	Source	Chordal	$13.1\pm 6.3$	$1.6\pm 1.5$
Source	Joint	Chordal	$\mathbf{10.5\pm 5.8}$	$\mathbf{1.1\pm 1.0}$
Target	Source	Chordal	$74.9\pm 52.8$	$5.2\pm 4.5$
	Target	Chordal	$33.4\pm 36.4$	$4.4\pm 4.5$
	Joint	Chordal	$16.0\pm 11.8$	$2.9\pm 2.5$

Table 5: Quantitative evaluation of reorientation models on held-out test sets from the source (free-running;

n=21

subjects) and target (Doppler-gated;

n=7

subjects) domains shown as mean

\pm

standard deviation. Each image was sampled

250

times using independently drawn random rigid transformations. The best performance within each domain is shown in bold (linear mixed-effect model comparison with Bonferroni correction,

p<0.05

). † Target domain GD inter-rater agreement was computed on a single reorientation per subject (

n=7

) and is not directly comparable to network performance assessed over whole range of rotations.

SPARC pipeline evaluation on clinical cohort ( $n=121)$
1) Manual intervention summary
Step	$n$ ( $\%$ )	Time per case
Manual stack exclusion	6 (5.0%)	$\sim$ 30 s
Manual template selection	3 (2.5%)	$\sim$ 60 s
Manual thoracic segmentation	3 (2.5%)	$\sim$ 120 s
Manual anatomical reorientation	12 (9.9%)	$\sim$ 60 s

Table 6: End-to-end SPARC pipeline evaluation on a held-out clinical cohort (

n=121

subjects). 1) Manual intervention summary: the number and estimated duration of manual interventions required at each pipeline stage. 2) Template stack selection evaluation. 3) Thoracic segmentation evaluation. 4) Slice-to-volume reconstruction evaluation with mean processing time. 5) Anatomical reorientation evaluation stratified by cardiac centre localisation ablation. 6) End-to-end pipeline summary: overall counts and mean processing time for fully automatic and semi-automatic runs.

Figure S2: Geodesic distance (GD) of the Transfer approach as a function of simulated rotation magnitude $|\theta|$ , evaluated on the held-out target domain test set (Doppler-gated; $n=7$ subjects, 250 randomly sampled rigid transformations per subject). Each point represents one simulated transformation, with subjects colour-coded.

Explore similar work

End-to-End 4D Heart Mesh Recovery Across Full-Stack and Sparse Cardiac MRI
Sep 15, 2025Yihong Chen, Jiancheng Yang, Deniz Sayin Mercadier +34D Reconstruction 3D Reconstruction
School of Computer and Communication Science, EPFL · ELLIS Institute Finland, Aalto University, Finland · University of North Carolina at Charlotte +1
MRI-Guided Reslice-Refined Cross-Slice SDF Reconstruction of the Left Ventricle from Cardiac MRI with Sparse Axial Supervision
Sep 8, 2026Quanxin Zheng, Shuai ZhaoImplicit Neural Representations Signed Distance Functions
1China Electronics (Beijing) Information Technology Research Institute Co., Ltd.
CineMesh4D: Personalized 4D Whole Heart Reconstruction from Sparse Cine MRI
May 13, 2026Xiaoyue Liu, Xiaohan Yuan, Mark Y Chan +23D Mesh Generation 4D Reconstruction
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

Sep 15, 2025J/K move · Enter open · S save

End-to-End 4D Heart Mesh Recovery Across Full-Stack and Sparse Cardiac MRI

Yihong Chen, Jiancheng Yang, Deniz Sayin Mercadier, Hieu Le+2

School of Computer and Communication Science, EPFL · ELLIS Institute Finland, Aalto University, Finland · University of North Carolina at Charlotte+1

4D Reconstruction 3D Reconstruction Image Reconstruction Sparse-View 3D Reconstruction Magnetic Resonance Imaging Cardiac MRI

Open paper AskPDF arXiv

Reconstructing cardiac motion from CMR sequences is critical for diagnosis, prognosis, and intervention. Existing methods rely on complete CMR stacks to infer full heart motion, limiting their applicability during intervention when only sparse observations are available. We present TetHeart, the first end-to-end framework for unified 4D heart mesh recovery from both offline full-stack and intra-procedural sparse-slice observations. Our method leverages deformable tetrahedra to capture shape and motion in a coherent space shared across cardiac structures. Before a procedure, it initializes detailed, patient-specific heart meshes from high-quality full stacks, which can then be updated using whatever slices can be obtained in real-time, down to a single one during the procedure. TetHeart incorporates several key innovations: (i) an attentive slice-adaptive 2D-3D feature assembly mechanism that integrates information from arbitrary numbers of slices at any position; (ii) a distillation strategy to ensure accurate reconstruction under extreme sparsity; and (iii) a weakly supervised motion learning scheme requiring annotations only at keyframes, such as the end-diastolic and end-systolic phases. Trained and validated on three large public datasets and evaluated zero-shot on additional private interventional and public datasets without retraining, TetHeart achieves state-of-the-art accuracy and strong generalization in both pre- and intra-procedural settings. Code and dataset is available at https://github.com/Scalsol/TetHeart.

Abstract

Explore similar work

End-to-End 4D Heart Mesh Recovery Across Full-Stack and Sparse Cardiac MRI

MRI-Guided Reslice-Refined Cross-Slice SDF Reconstruction of the Left Ventricle from Cardiac MRI with Sparse Axial Supervision

CineMesh4D: Personalized 4D Whole Heart Reconstruction from Sparse Cine MRI