Automatic segmentation of isotropic free-running 4D whole-heart CMR is more reproducible and accurate than that of 2D cine

Presenting Author(s)

• AO

  Augustin C. Ogier, PhD

  Post-doctoral fellow
  Lausanne University Hospital (CHUV) and University of Lausanne (UNIL), Switzerland

Primary Author(s)

• AO

  Augustin C. Ogier, PhD

  Post-doctoral fellow
  Lausanne University Hospital (CHUV) and University of Lausanne (UNIL), Switzerland

Co-Author(s)

• SB

  Salomé Baup, MSc

  Master student
  Lausanne University Hospital (CHUV) and University of Lausanne (UNIL), Switzerland
• GI

  Gorun Ilanjian, MSc

  Médecin assistant
  Lausanne University Hospital (CHUV) and University of Lausanne (UNIL), Switzerland
• AR

  Angela Rocca, MSc

  PhD student
  Lausanne University Hospital (CHUV) and University of Lausanne (UNIL), Switzerland
• JB

  Jaume Banus, PhD

  Post-doctoral fellow
  Lausanne University Hospital (CHUV) and University of Lausanne (UNIL), Switzerland
• MN

  Martin Nicoletti, MSc

  PhD student
  Lausanne University Hospital (CHUV) and University of Lausanne (UNIL), Switzerland
• JL

  Jean-Baptiste Ledoux, MSc

  Senior Technologist
  Lausanne University Hospital (CHUV), Switzerland
• JR

  Jonas Richiardi, PhD

  Principal investigator
  University Hospital (CHUV) and University of Lausanne (UNIL), Switzerland
• 

  Mathias Stuber, PhD

  Full Professor
  CIBM-CHUV-UNIL
  Laussane University, Switzerland
• DR

  David Rotzinger, MD

  Associate radiologist
  Lausanne University Hospital (CHUV) and University of Lausanne (UNIL), Switzerland
• RH

  Roger Hullin, MD

  Chief Severe Heart Failure en Heart Transplantation Program
  Lausanne University Hospital (CHUV), Switzerland
• 

  Ruud B. Van Heeswijk, PhD

  Senior Lecturer
  Lausanne University Hospital (CHUV), Switzerland

Background: Free-running (FR) whole-heart methods can simplify cardiac MRI workflow, since they do not need ECG gating or breath-holding, while the resulting 5D (isotropic 3D + cardiac cycle + respiratory phases) images can be reformatted in any desired orientation a posteriori¹. Moreover, this isotropic spatial resolution should improve the accuracy of manual segmentation compared to routine 2D cine imaging. However, this has never been validated, since the wealth of information contained in 5D images (easily 6’000 images) hinders full manual segmentation and broader clinical adoption. This study addresses this challenge by designing, testing, and validating a deep learning (DL) framework for the automatic reproducible and precise segmentation of isotropic 3D whole-heart dynamic (4D) FR images.


Methods: Twenty participants (10 volunteers and 10 heart failure patients, 63 ± 13 years, 6 F) were scanned on a 3T scanner (PrismaFit, Siemens) with a FR3D radial¹ gradient-echo sequence (isotropic voxel size = 1.37 mm, field of view = 220mm). Following a bolus injection of 0.1 mL/kg gadobutrol, self-gating and compressed sensing were used to reconstruct 3D images for each cardiac and respiratory phase^1,2. End-expiratory images were selected for analysis, resulting in 4D datasets. As a reference, 2D cine images were also acquired. Manual segmentation of the left-ventricular blood pool (LVB) and myocardium (LVM) was performed on cine images (𝐶𝑖𝑛𝑒_𝑀) and semi-automatically (i.e. manually with slice-to-slice propagation)³ on 4D FR images (𝐹𝑅_𝑀), both at end-systole and end-diastole. DL training was conducted on 𝐹𝑅_𝑀 using nnU-Net⁴, with 5- fold cross-validation. The accuracy of the DL-based automatic segmentation (𝐹𝑅_𝐴) was evaluated using the Dice similarity coefficient (DSC), and the resulting ejection fraction (EF) compared via Bland-Altman analyses and intraclass correlation coefficients via ICC(2,1). The consistency error of LVM volumes between diastole and systole was compared across all segmentation types. A second expert similarly segmented all images from a subset of 10 subjects to assess inter-observer variability.


Results: The 𝐹𝑅_𝐴 segmentation (Fig. 1) demonstrates high accuracy compared to 𝐹𝑅_𝑀, with DSC values of 0.87 ± 0.02 for LVM and 0.94 ± 0.01 for LVB (Fig. 2A). The EF measured with 𝐹𝑅_𝐴 and 𝐹𝑅_𝑀 agreed strongly (ICC = 0.99, Fig .2B), as did the EF measured with 𝐶𝑖𝑛𝑒_𝑀 and 𝐹𝑅_𝑀 (ICC = 0.95, Fig. 2C), though with wider limits of agreement. The LVM volume consistency error (Fig. 3A) was highest for 𝐶𝑖𝑛𝑒_𝑀 (11.0 ± 8.5%), followed by 𝐹𝑅_𝑀 (8.5 ± 5.4%), and lowest for 𝐹𝑅_𝐴 (6.9 ± 3.2%). Inter-observer agreement was comparable for 𝐹𝑅_𝑀 (ICC = 0.98, Fig. 3B) and 𝐶𝑖𝑛𝑒_𝑀 (ICC = 0.97, Fig. 3C) segmentations, with 𝐹𝑅_𝑀 exhibiting slightly greater bias but narrower limits of agreement.

Conclusion: We successfully validated a DL framework for the accurate and reproducible segmentation of spatially isotropic 4D whole-heart images. The high DSC, strong agreement in EF (comparable to intra-expert variability of the literature⁵) and lower LVM volume consistency error across cardiac phases underscore the potential of FR imaging to surpass 2D cine metrics. These more reproducible and consistent cardiac function measurements thus pave the way for broader clinical adoption of FR 5D cardiac MRI. Future efforts will focus on expanding validation to different image contrasts and field strengths, towards implementing push-button high-precision cardiac MRI.

DL-based automatic segmentation (FR_A) of the spatially isotropic 4D FR images at end-systole (left) and end-diastole (right) for a healthy volunteer, a heart failure with preserved ejection fraction (HFpEF) patient and a heart failure with reduced ejection fraction (HFrEF) patient. FR images are displayed in pseudo two-chamber, pseudo four-chamber, and mid-ventricular short-axis views, with the three planes oriented perpendicular to each other as used in the semi-automatic segmentation that served as the ground truth for the DL-based training.

Comparison between the semi-automatic segmentation (FR_M), the DL-based automatic segmentation (FR_A) and the manual segmentation of the cine images (Cine_M). A: Boxplots of the Dice similarity coefficients (DSC) on the FR images for the left-ventricular myocardium (LVM) and the left-ventricular blood pool (LVB) segmentation. B-C: Bland-Altman analyses of the left-ventricular ejection fraction computed from the Cine_M segmentation, the FR_M segmentation and the FR_A segmentation.

Segmentation reproducibility. A: Consistency error of the left-ventricular (LV) myocardium volume was computed as the difference between the end-diastole and end-systole volume (∆_LVM) divided by the mean volume (μ_LVM) for the manual segmentation of the Cine images (Cine_M), the semi-automatic segmentation (FR_M) and the DL-based automatic segmentation (FR_A) of the FR images on 20 subjects. B-C: On a subset of 10 subjects, a second expert segmented in a similar way both Cine images (Cine_M2) and FR images (FR_M2). B: Bland-Altman analysis between the LV ejection fraction from the manual segmentation of the two experts on cine images. C: Bland-Altman analysis between the LV ejection fraction from the semi-automatic segmentation of the two experts on FR images.