MorphiNet: A Graph Subdivision Network for Adaptive Bi-ventricle Surface Reconstruction

TL;DR

MorphiNet tackles the core challenge of reconstructing accurate bi-ventricular heart surfaces from anisotropic CMR data by learning cardiac anatomy from unpaired CT data and transferring it to CMR through a gradient-field deformation of a fixed template mesh. It integrates dynUNet-based segmentation, ResNet-based complementary segmentation, a gradient field derived from a distance map, and an adaptive Graph Subdivision Network to refine meshes while preserving dense point correspondence. Compared with state-of-the-art template-based methods, MorphiNet delivers higher anatomical fidelity with Dice near $0.8$ on CT datasets and substantially faster inference than neural implicit approaches (≈$2.7$ s vs. ≈$133$ s), with robust zero-shot generalization to CAP/ACDC data (Dice ≈ $0.7$). Motion tracking demonstrates functional realism with ejection fractions in the range of current clinical benchmarks, enabling reliable cardiac function analysis while admitting trade-offs in mesh quality near thin regions like valve planes.

Abstract

Cardiac Magnetic Resonance (CMR) imaging is widely used for heart model reconstruction and digital twin computational analysis because of its ability to visualize soft tissues and capture dynamic functions. However, CMR images have an anisotropic nature, characterized by large inter-slice distances and misalignments from cardiac motion. These limitations result in data loss and measurement inaccuracies, hindering the capture of detailed anatomical structures. In this work, we introduce MorphiNet, a novel network that reproduces heart anatomy learned from high-resolution Computed Tomography (CT) images, unpaired with CMR images. MorphiNet encodes the anatomical structure as gradient fields, deforming template meshes into patient-specific geometries. A multilayer graph subdivision network refines these geometries while maintaining a dense point correspondence, suitable for computational analysis. MorphiNet achieved state-of-the-art bi-ventricular myocardium reconstruction on CMR patients with tetralogy of Fallot with 0.3 higher Dice score and 2.6 lower Hausdorff distance compared to the best existing template-based methods. While matching the anatomical fidelity of comparable neural implicit function methods, MorphiNet delivered 50$\times$ faster inference. Cross-dataset validation on the Automated Cardiac Diagnosis Challenge confirmed robust generalization, achieving a 0.7 Dice score with 30\% improvement over previous template-based approaches. We validate our anatomical learning approach through the successful restoration of missing cardiac structures and demonstrate significant improvement over standard Loop subdivision. Motion tracking experiments further confirm MorphiNet's capability for cardiac function analysis, including accurate ejection fraction calculation that correctly identifies myocardial dysfunction in tetralogy of Fallot patients.

Figures (8)

• Figure 1: Overview of the proposed MorphiNet framework. From top to down: a) Segment the cardiac magnetic resonance images with large inter-slice distances and misalignments. b) Complete specific segmentation. c) Find a gradient field from the segmentation to adjust the template mesh. d) End with a reconstructed heart model preserving underlying cardiac anatomy while conforming to patient-specific data.
• Figure 2: Diagram showing MorphiNet's training and inference workflows. MorphiNet consists of three neural network modules with learnable parameters: dynUNet, ResNet, and GSN. Training phase: CT images follow CT dynUNet$\ \rightarrow\ $ResNet$\ \rightarrow\ $GSN pathway; CMR images follow CMR dynUNet pathway only. Inference phase: CMR images are processed through CMR dynUNet$\ \rightarrow\ $ResNet$\ \rightarrow\ $GSN when model parameters are frozen. CT inference follows CT dynUNet$\ \rightarrow\ $ResNet$\ \rightarrow\ $GSN pathway.
• Figure 3: The patient-specific adjustment applied to the template model. (a) A rigid "swing" rotation is determined to register the RV centroid in the template mesh with the RV centroid in the distance map (shown as extracted myocardium surface in the figures for ease of demonstration). (b) The gradient field is calculated from the distance map and is used to deform the template mesh iteratively along the reverse direction of the gradient vectors. The gradient field deformation is applied individually to groups of the template mesh's vertices categorized by their anatomical label. Only LV and RV epicardium are shown in the figures for ease of demonstration.
• Figure 4: Surface reconstruction results on SCOT-HEART and MMWHS CT datasets, with the highest face count per method displayed in the bottom left and colored with surface errors (mm). The RV myocardium segmentation in the MMWHS dataset grows from RV epicardium, resulting in a 3 mm-thick myocardium free-wall segmentation.
• Figure 5: Surface reconstruction results on the CAP and ACDC CMR datasets, depicting myocardium mesh slices in three views: 1) short-axis, 2) three-chamber, and 3) four-chamber. Note that the ACDC dataset provides no LAX image-view data. Those image-view planes are labeled as color boxes and positioned against respective reconstructed models. In each boxed image view, the model's cross-section (blue) is visualized in contrast with the manual segmentation (red). The RV myocardium segmentation in the ACDC dataset grows from RV epicardium, resulting in a 3 mm-thick myocardium free-wall segmentation.