Deep-Learning-Based Segmentation of Gridded Reconstructions from Undersampled Free-Running 4D Data for On-Scanner Deployment

Presenting Author(s)

• 

  Augustin C. Ogier, PhD

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

Primary Author(s)

• 

  Augustin C. Ogier, PhD

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

Co-Author(s)

• GP

  Gabriel Paffi, MSc

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

  Jérôme Yerly, PhD

  Senior scientist
  Lausanne University Hospital (CHUV) and University of Lausanne (UNIL)
  Lausanne, Vaud, Switzerland
• AR

  Angela Rocca, MSc

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

  Isabel Montón Quesada, MSc

  PhD student
  Lausanne University Hospital (CHUV)
  Lausanne, Vaud, Switzerland
• JL

  Jean-Baptiste Ledoux, MSc

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

  Robert J. Holtackers, PhD

  MR Physicist & Assistant Professor
  Maastricht University Medical Centre
  Maastricht, Limburg, Netherlands
• MS

  Matthias Stuber, PhD

  Professor/Director
  CIBM/CHUV/UNIL
  Lausanne, Switzerland
• 

  Christopher W. Roy, PhD

  Lecturer
  University Hospital (CHUV) and University of Lausanne (UNIL)
  Lausanne, Vaud, Switzerland
• RH

  Roger Hullin, MD

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

  David Rotzinger, MD

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

  Ruud B. van Heeswijk, PhD, FSCMR

  Senior lecturer
  Lausanne University Hospital (CHUV) and University of Lausanne (UNIL)
  Lausanne, Vaud, Switzerland

Background: Free-running (FR) cardiac MRI enables free-breathing, ECG-free, 5D cardiac imaging (3D spatial + cardiac + respiratory).¹ However, the wealth of information renders manual segmentation impractical, highlighting the need for automated analysis approaches. Currently, automatic segmentation follows time-consuming compressed-sensing (CS) image reconstruction,² limiting the analysis to retrospective, off-scanner workflows.³ Here, we extend a previously proposed deep learning (DL) model to directly segment computationally efficient gridded image reconstructions without CS, enabling direct on-scanner use.

Methods: 3D radial FR data were acquired in 35 subjects: GRE in 10 healthy volunteers and 10 heart failure patients at 3T (PrismaFit, Siemens) post-gadobutrol,⁴ and contrast-agent-free bSSFP in 15 healthy volunteers at 1.5 T (Sola, Siemens).⁵ Data were acquired with a 1.38 mm isotropic resolution. Cardiac and respiratory self-gating signals were used to sort the data into 4 respiratory bins and 50 ms cardiac frames, yielding highly undersampled 5D k-space bins. Three reconstruction methods were investigated: gridded, CS, and fNav. Gridded and CS produced 5D cardiac- and respiratory-motion–resolved images (only end-expiration was used in this study), whereas fNav generated 4D cardiac-motion–resolved, respiratory-motion–corrected images.⁶ Gridded and fNav images can be generated within minutes but may suffer in quality due to undersampling, while CS typically requires hours to produce higher-quality images.¹ Leveraging spatial alignment across reconstructions, semi-manual segmentation⁷ of the ventricular cavities and myocardium was performed only on CS images, and the same labels were used to train nnUNet⁸ models separately on gridded, fNav, and CS images in a 5-fold cross validation scheme.³ Performance was assessed with the Dice similarity coefficient (DSC), relative volume difference (RVD), ejection fraction (EF) agreement, and myocardial volume mismatch across cardiac phases.

Results: All three automatic segmentations closely matched ground-truth (Fig. 1) with strong geometric accuracy (Table 1). They also demonstrated strong ejection fraction agreement, with negligible bias ( < 1 mL), narrow limits of agreement, and high ICC ( >0.96) (Fig. 2A–C). Myocardial volume was conserved across cardiac phases in all automatic segmentations and did not differ significantly from the ground-truth segmentation (Fig. 2D).

Conclusion: We demonstrated that the proposed DL segmentation models deliver clinically equivalent performance on rapidly reconstructed gridded and fNav images compared to the high‐quality CS reference. These results indicate that CS is not necessary for accurate whole-heart delineation, enabling immediate post-acquisition analysis. The models have already been implemented for on-scanner deployment. Future prospective studies should quantify their impact on workflow efficiency and diagnostic accuracy to pave the way for FR whole-heart cardiac MRI in clinical routine.

Table 1: Quantitative segmentation performance across reconstruction models. Dice similarity coefficient (DSC) and relative volume difference (RVD) are reported for the left ventricular blood pool (LVB), left ventricular myocardium (LVM), and right ventricular blood pool (RVB). Values are presented as mean ± standard deviation, comparing DL-based automatic segmentation trained on compressed-sensing images (CS-FRA), respiratory motion corrected images (fNav-FRA), and gridded reconstructed images (Gridded-FRA) against the ground truth segmentation performed on compressed-sensing images.

 

Metric	Region	Model
		CS-FRA	fNav-FRA	Gridded-FRA
DSC (1)	LVB	0.94 ± 0.01	0.93 ± 0.02	0.93 ± 0.01
	LVM	0.86 ± 0.02	0.82 ± 0.04	0.82 ± 0.04
	RVB	0.92 ± 0.02	0.90 ± 0.02	0.91 ± 0.02
RVD (%)	LVB	2.4 ± 2.1	3.1 ± 2.4	4.1 ± 2.5
	LVM	5.8 ± 3.7	6.6 ± 4.3	7.0 ± 5.3
	RVB	4.5 ± 2.8	5.6 ± 4.4	5.2 ± 3.8

Figure 1: Visual comparison of manual and DL-based automatic segmentations on spatially isotropic 4D free-running cardiac MRI at end-diastole (top) and end-systole (bottom) in a healthy volunteer in their 60s. Short-axis and pseudo four-chamber views are shown for each phase. The reference semi-automatic segmentation (CS-FRM) was performed on compressed sensing (CS) images, while automatic segmentations were generated using a deep learning model trained on CS images (CS-FRA), respiratory motion-corrected images (fNav-FRA), and gridded reconstructed images (Gridded-FRA). For each reconstruction method, images are shown both without and with segmentation overlays. Segmented structures include the left ventricular myocardium (green), left ventricular blood pool (red), and right ventricular blood pool (purple).
Figure 2: Bland-Altman analysis of left-ventricular ejection fraction comparing semi-manual segmentation on compressed-sensing images (CS-FRM) with the DL-based automatic segmentations trained and applied to A) compressed-sensing images (CS-FRA), B) respiratory motion corrected images (fNav-FRA) and C) gridded images (Gridded-FRA). D) Systolic–diastolic left ventricular myocardial (LVM) volume mismatch computed for each segmentation as the difference between the end-diastole and end-systole volumes (∆_LVM) divided by the mean volume (μ_LVM).