Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

AI-enabled cardiac shape reconstruction from routine magnetic resonance imaging

Tanmay Mukherjee, Neil Gautam, Nikhil Kadivar, Elizabeth M. Fugate, Kyle J. Myers, Diana Lindquist, Pierre Croisille, Sakthivel Sadayappan, Patrick Clarysse, Jacques Ohayon, Roderic Pettigrew, George Karniadakis, Reza Avazmohammadi

Abstract

Computational models of cardiac structure and function are increasingly central to the development of subject-specific cardiac digital twins, enabling improved characterization of contractile dysfunction, pathological remodeling, and electrical abnormalities. A critical prerequisite for these models is the accurate reconstruction of three-dimensional (3D) cardiac anatomy from medical imaging. Multi-planar magnetic resonance imaging, particularly when combined with artificial intelligence, offers a clinically feasible alternative to conventional reconstruction techniques. In this study, we present a neural field-based reconstruction framework that recovers 3D cardiac geometries from sparse planar contour data by learning continuous shape representations. Reconstruction performance was evaluated using complementary in-silico and in vivo datasets spanning variations in sampling density and geometric complexity. Across both datasets, reconstructed meshes closely matched reference geometries, demonstrating that the neural field approach faithfully captures cardiac planar contours. Compared with traditional local interpolation methods, the proposed framework exhibited improved geometric fidelity in anatomically challenging regions, including the left ventricular apex and basal segments, particularly under sparse sampling conditions. Collectively, these findings demonstrate that neural field-based reconstruction provides a robust and efficient pathway for multi-planar cardiac shape recovery, with particular relevance for AI-driven modeling pipelines and data-limited settings such as small-animal and time-resolved cardiac imaging.

AI-enabled cardiac shape reconstruction from routine magnetic resonance imaging

Abstract

Computational models of cardiac structure and function are increasingly central to the development of subject-specific cardiac digital twins, enabling improved characterization of contractile dysfunction, pathological remodeling, and electrical abnormalities. A critical prerequisite for these models is the accurate reconstruction of three-dimensional (3D) cardiac anatomy from medical imaging. Multi-planar magnetic resonance imaging, particularly when combined with artificial intelligence, offers a clinically feasible alternative to conventional reconstruction techniques. In this study, we present a neural field-based reconstruction framework that recovers 3D cardiac geometries from sparse planar contour data by learning continuous shape representations. Reconstruction performance was evaluated using complementary in-silico and in vivo datasets spanning variations in sampling density and geometric complexity. Across both datasets, reconstructed meshes closely matched reference geometries, demonstrating that the neural field approach faithfully captures cardiac planar contours. Compared with traditional local interpolation methods, the proposed framework exhibited improved geometric fidelity in anatomically challenging regions, including the left ventricular apex and basal segments, particularly under sparse sampling conditions. Collectively, these findings demonstrate that neural field-based reconstruction provides a robust and efficient pathway for multi-planar cardiac shape recovery, with particular relevance for AI-driven modeling pipelines and data-limited settings such as small-animal and time-resolved cardiac imaging.

Paper Structure

This paper contains 26 sections, 7 equations, 11 figures, 1 table.

Table of Contents

  1. Introduction
  2. Related Work
  3. Traditional learning-based methods
  4. Neural fields for anatomical shape reconstruction
  5. Methods
  6. Data acquisition and preprocessing
  7. FE simulations
  8. Multi-view MRI
  9. Dataset sampling
  10. Neural field-based reconstruction
  11. Evaluation of spatial reconstruction
  12. Geometric correspondence analysis
  13. Curvature analysis
  14. Results
  15. Evolution of training loss
  16. ...and 11 more sections

Figures (11)

  • Figure 1: Flowchart describing the workflow for cardiac shape reconstruction from in-plane segmentation contours. (A) Multi-view magnetic resonance imaging (MRI) and in-plane segmentation of the datasets (mouse and human-specific) (B) Ground-truth reconstruction at end-diastole (ED) and end-systole (ES), with finite element simulations of the to augment the dataset by creating several geometries between ED and ES, (C) hierarchical sampling of points inside, outside, and on the boundary to create a geometric representation of the cardiac shape, (D) learning the input representation using a neural field (NF), (E) Refining the NF iteratively to increase the sampling of points describing the cardiac shape, (F) reconstruction through a meshing algorithm, and evaluation against ground-truth geometries.
  • Figure 2: Evolution of training loss for the neural field for cardiac reconstruction in the finite element (biventricular mouse geometries) and in vivo (single ventricle mouse and human) datasets. Training losses are presented for (A) binary cross-entropy loss measuring the accuracy of the inside/outside indicator function, and (B) Eikonal hinge loss promoting signed distance function properties and surface smoothness. Each level of refinement (0-5) consists of 120 iterations.
  • Figure 3: Comparison of ground-truth and reconstructed contours at multiple short- and long-axis (SAX and LAX) planes for FE dataset. (A) Location of SAX and LAX planes used for the comparison, highlighting the contours at the LV endocardium. (B) Error map of the closest distance between points in both sets of endocardial contours. (C and D) Histogram detailing the distribution of errors at different points in SAX and LAX planes, respectively. Points were normalized prior to quantitative comparison.
  • Figure 4: Comparison of ground-truth and reconstructed contours of the right ventricular endocardium at multiple short- and long-axis (SAX and LAX) planes for the FE dataset. (A and B) Histogram detailing the distribution of bidirectional chamfer distance at different points in SAX and LAX planes, respectively. Points were normalized prior to quantitative comprison.
  • Figure 5: Comparison of ground-truth and reconstructed contours of the left ventricular endocardium at different refinement levels for the FE dataset. (A) Heatmaps highlighting the bidirectional distance between the reconstrcuted LV endocardium at a specific refinement level versus the ground-truth. (B and C) Histogram of the heatmaps showing the relative changes in the distance-based error metric with each level at end-diastole and end-systole, respectively. Increasing levels correspond to a greater number of points used to represent the ground-truth geometry for neural field interpolation.
  • ...and 6 more figures