Comparison of semi-automatic myocardial scar segmentation methods on black-blood late gadolinium enhancement images

Presenting Author(s)

• 

  Thaïs Genisson, MSc

  PhD student
  IHU Liryc, Université de Bordeaux
  Bordeaux, Aquitaine, France

Primary Author(s)

• 

  Thaïs Genisson, MSc

  PhD student
  IHU Liryc, Université de Bordeaux
  Bordeaux, Aquitaine, France

Co-Author(s)

• 

  Kalvin Narceau, MSc

  Phd student
  Bordeaux University - INSERM U1045
  Pessac, Aquitaine, France
• TR

  Théo Richard, MSc

  Engineer
  IHU LIRYC, Heart rhythm disease institute, Université de Bordeaux – INSERM U1045, Avenue du Haut Lévêque, 33604, Pessac, France
  Bordeaux, Aquitaine, France
• 

  Victor de Villedon de Naide, MSc

  PhD student
  Bordeaux University - INSERM U1045
  Bordeaux, Aquitaine, France
• EB

  Ewan Barel

  Research engineer
  IHU Liryc, Université de Bordeaux, France
• EG

  Edouard Gerbaud, MD, PhD

  Cardiac Intensive Care Unit, Groupe Hospitalier Sud, CHU de Bordeaux, Pessac, France, France
• PJ

  Pierre Jaïs, MD, PhD

  PROF/PhD
  Hôpital Cardiologique du Haut-Lévêque, CHU de Bordeaux
  Bordeaux, Aquitaine, France
• MS

  Matthias Stuber, PhD

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

  Hubert Cochet, MD, PhD

  Professor
  Bordeaux University - INSERM U1045
  Bordeaux, Aquitaine, France
• 

  Aurelien Bustin, FSCMR

  Research Associate
  Department of Cardiovascular Imaging, Hôpital Cardiologique du Haut-Lévêque, CHU de Bordeaux, Avenue de Magellan, Pessac, France; IHU LIRYC, Electrophysiology and Heart Modeling Institute, Université de Bordeaux – INSERM U1045, Avenue du Haut Lévêque, Pessac, France; Department of Diagnostic and Interventional Radiology, Lausanne University Hospital and University of Lausanne, Lausanne, Switzerland
  Bordeaux, Aquitaine, France

Background: Myocardial scar assessment is key for patient prognosis with left ventricular (LV) scar size being a major predictor of cardiac events risks¹. Black-blood (BlB) late gadolinium enhancement (LGE) imaging improves scar visualization by nulling both blood and myocardium^2,3, solving the blood–scar interface contrast issue of bright-blood (BrB) LGE⁴. However, scar segmentation on BlB images remains manual, which is time-consuming and operator dependent.
We aim at assessing the performances of semi-automatic scar segmentation methods on BlB images.


Methods: The study included 20 patients (85% male, age range 49-75yo) with ischemic heart disease. Breath-held short-axis 2D whole-heart joint BrB and BlB SPOT⁵ images were collected on a 1.5T system (Siemens Aera) in mid-diastole. Resolution was 1.5x1.5mm² with 8mm slice thickness. Reference breath-held BrB PSIR⁴ images were collected with similar imaging parameters.

LV wall was manually segmented on the BrB SPOT images. Myocardial scar was manually segmented on the BlB images by two experts to assess inter-observer variability, and twice by one expert to assess intra-observer variability.Several semi-automatic scar segmentation techniques were evaluated:: n-Standard Deviation (n-SD, n from 2 to 15), and Full Width at Half Maximum (FWHM), 25% and 50% maximum signal intensity (MSI)⁶, and Otsu thresholding⁷.

Scar segmentation accuracy was assessed using the Dice Similarity Coefficient (DSC). Scar burden (LV wall %) was computed from the segmentation masks. Absolute differences in scar burden between manual and semi-automated methods were calculated. Intra and inter-observer variability in manual scar burden measurements were also computed. Agreement between manual and semi-automated scar burden measurements was assessed using Bland–Altman analysis for all the semi-automatic methods.

Results: For the n-SD method, the highest mean DSC across all patients was obtained with n=12 (73.5%). FWHM achieved a mean DSC of 64.9%, while the 50% MSI outperformed the 25% threshold (71.6% vs. 51.7%). Other methods performed poorly (Fig.1A).

Intra- and inter-observer mean of differences in scar burden was of 2.4% and 2.8%, respectively. For the n-SD method, the smallest mean scar burden difference was obtained with n=14 (7.3%). FWHM achieved a mean scar burden difference of 12.0%, while 50% MSI obtained a mean difference of 11.8% (Fig.1B).

Among the best-performing methods (11SD, 12SD, 13SD, 14SD, 50% MSI, and FWHM), Bland–Altman analysis showed that 13SD and FWHM provided the smallest biases and narrowest limits of agreement (+0.0±19.8% and –0.7±29%, respectively, Fig.1C&1D). These two methods also showed good visual agreement with the manual reference (Figs.2&3).

Conclusion: This study shows that semi-automatic scar segmentation methods used on BrB images can be applied to BlB images. The n-SD method is most robust for n between 10 and 15, while FWHM and the 50% MSI threshold also allow fast and reliable scar assessment for clinical use.

Fig.1: A) Table reporting the mean Dice Similarity Coefficient (DSC), averaged over all patients, for all semi-automatic segmentation methods. B) Box-plot of the absolute difference in scar burden for all semi-automatic segmentation methods C) Bland-Atlman analysis comparing scar burden between manual and 13-SD scar segmentation method D) Bland-Atlman analysis comparing scar burden between manual and FWHM scar segmentation method.
Fig.2: Conventional PSIR images, SPOT images, and scar segmentation masks obtained from black-blood images using manual, 13SD, and FWHM methods in three patients with ischemic heart disease.
Fig.3: Full ventricular view of SPOT images and scar segmentation for manual, 13-SD and FWHM methods obtained in a patient 51-years-old-male with myocardial infarction originating from the right coronary artery.