Advanced myocardial scar segmentation with automated analysis of joint bright- and black-blood late gadolinium enhancement images

Friday, January 31, 2025

11:50 AM – 12:00 PM East Coast USA Time

Location: Blue Room

Presenting Author(s)

TG

Thaïs Génisson, MSc

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

Primary Author(s)

TG

Thaïs Génisson, MSc

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

Co-Author(s)

Vd

Victor de Villedon de Naide, MSc

PhD student
Bordeaux University - INSERM U1045, France
KN

Kalvin Narceau, MSc

Research Engineer
Bordeaux University - INSERM U1045, France
bd

baptiste durand, MD, PhD

Radiologist
Department of Cardiovascular Imaging, Hôpital Cardiologique du Haut-Lévêque, CHU de Bordeaux, Avenue de Magellan, 33604 Pessac, France, France
JM

Jean-David Maes, MD

PhD student
Bordeaux University - INSERM U1045, France
MV

Manuel Villegas-Martinez, PhD

Postdoctoral Student
Bordeaux University - INSERM U1045, France
TR

Théo Richard, MSc

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

Kun He, PhD, MSc

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

Pauline Gut, MSc

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

Pierre Jaïs, MD, PhD

PROF/PhD
Hôpital Cardiologique du Haut-Lévêque, CHU de Bordeaux, France
Mathias Stuber, PhD

Full Professor
CIBM-CHUV-UNIL
Laussane University, Switzerland
HC

Hubert Cochet, MD, PhD

Professor
Department of Cardiovascular Imaging, Hôpital Cardiologique du Haut-Lévêque, CHU de Bordeaux, France
Aurelien Bustin, PhD

Assistant Professor
IHU LIRYC, France

Background:

Assessing myocardial scar characteristics on magnetic resonance imaging (MRI) is crucial for patient prognostic and effective therapy, with left ventricle (LV) scar size being a key predictor of risk for cardiac events¹. However, the poor scar-to-blood contrast in clinically-used bright-blood late gadolinium enhancement (LGE) images limits the robustness of automated segmentation algorithms for quantifying myocardial scars^2,3. As a result, highly trained radiologists must manually segment the scars, a process that is time-consuming, operator-dependent and labour-intensive. In this study, we propose a post-processing solution leveraging artificial intelligence to automate scar segmentation for improved analysis of joint bright- and black-blood LGE images (SPOT sequence⁴).

Methods:

Data A dataset composed of 70 patients (21% female, age range 28-81yo) with known or suspected ischemic heart disease was divided into a training (70%), a validation (10%) and a testing (20%) set. Breath-held short-axis 2D whole-heart SPOT images were collected on a 1.5T Siemens Aera system. For each slice, 2 dummy heartbeats and 8 single-shot images (5 bright-blood and 5 black-blood images) were acquired in mid-diastole. The image resolution was 1.5 mm x 1.5 mm with an 8 mm slice thickness. Reference breath-held phase-sensitive inversion recovery PSIR⁵ images were also collected with similar imaging parameters.

Processing The pipeline involved three steps (Fig.1):

i) Automated LV segmentation on bright-blood images using a transformer-based model⁶,

ii) Propagation of the LV contours onto the black-blood images and exclusion of data outside these contours,

iii) Segmentation of the LV scar on the preprocessed black-blood images using a U-net model⁷.

Analysis The scar segmentation accuracy was assessed using the Dice Similarity Coefficient (DSC). An experienced radiologist evaluated the U-net model segmentations using a Likert scale (0-3): 0=redo, 1=major adjustments, 2=minor adjustments, 3=no adjustments needed. The obtained scar size, as a proportion of the LV, was computed from the segmentation masks. The agreement between manual and automatic scar size and between SPOT and PSIR images was evaluated with Bland-Altman analysis (expressed as bias ±1.96 standard deviation). Processing times were recorded.

Results: The proposed algorithm segmented the scar in about 0.14 seconds per slice, achieving a mean DSC of 76.1% on the test set (Fig.2). The Likert score confirmed these results, with 62% of the predicted scar segmentations score of 3, 33% requiring some level of adjustment and only 5% requiring revision (Fig.3A). A good agreement in scar size was found between automatic and manual segmentations on SPOT with minimal bias (+4.63 % ± 14.9, Fig.3B). There was good agreement between the manual segmentation on reference PSIR images and the automated method using SPOT images (bias +9.54 % ± 20.88, Fig.3C).

Conclusion:
The proposed method for automated scar segmentation on bright- and black-blood LGE images achieved accurate results, closely matching manual segmentation on both SPOT and reference PSIR images.

Fig.1: Diagram of the proposed pipeline for automated myocardial scar segmentation. Left: SPOT collects 2D co-registered bright-blood and black-blood late gadolinium enhancement images in a single sequence. Middle: A transformer-based model is employed to segment the left ventricular wall fully automatically on bright-blood SPOT images and the pixels inside the wall are extracted. The left ventricle contours are propagated onto the black-blood SPOT image. Right: A U-net model is used to automatically segment the myocardial scars on the latter. Scar contours are propagated onto the bright-blood images to visualize the scar location and scar burden.

Fig.2: Full ventricular view of SPOT images and scar segmentation prediction obtained in a 63-year-old-male patient with myocardial infarction originating from the left anterior descending artery.

Fig.3: A) Proportion of test images by Likert scale values. B) Bland-Altman analysis for scar size values as a percentage of the left ventricle comparing manual and automatic scar segmentation on SPOT images for all patients in the test set. C) Bland-Altman analysis for scar size values as a percentage of the left ventricle comparing manual scar segmentation on PSIR reference images and automatic scar segmentation on SPOT images for all patients in the test set.