Rapid, Efficient Imaging
Aurélien Bustin, PhD
Junior Professor
IHU LIRYC
Bordeaux, Aquitaine, France
Aurélien Bustin, PhD
Junior Professor
IHU LIRYC
Bordeaux, Aquitaine, France
Indra Ribal, MSc
MSc
IHU Liryc, Université de Bordeaux, France
Géraldine Montier, MD
MD
Centre Hospitalier Universitaire de Bordeaux, France
Jean-David Maes, MD
MD
Centre Hospitalier Universitaire de Bordeaux, France
Thibault Boullé, MD
MD
Centre Hospitalier Universitaire de Bordeaux, France
Victor de Villedon de Naide, MSc
MSc
IHU LIRYC, Electrophysiology and Heart Modeling Institute, Université de Bordeaux – INSERM U1045, France
Pauline Gut, MSc
MSc
University Hospital (CHUV) and University of Lausanne (UNIL), Switzerland
Soumaya Sridi, MD
MD
Centre Hospitalier Universitaire de Bordeaux, France
Valery Ozenne, PhD
Dr
CNRS, Aquitaine, France
Marta Nuñez-Garcia, PhD
Dr
IHU LIRYC, Electrophysiology and Heart Modeling Institute, Université de Bordeaux – INSERM U1045, France
Maxime Sermesant, PhD
PROF/PhD
IHU LIRYC, Electrophysiology and Heart Modeling Institute, Université de Bordeaux – INSERM U1045, France
Michel Montaudon, MD, PhD
PROF/PhD
Hôpital Cardiologique du Haut-Lévêque, CHU de Bordeaux
Pessac, Aquitaine, France
Gäel Dournes, MD, PhD
PROF/PhD
Hôpital Cardiologique du Haut-Lévêque, CHU de Bordeaux, France
François Laurent, MD, PhD
PROF/PhD
Hôpital Cardiologique du Haut-Lévêque, CHU de Bordeaux
Pessac, Aquitaine, France
Pierre Jaïs, MD, PhD
PROF/PhD
Hôpital Cardiologique du Haut-Lévêque, CHU de Bordeaux
Bordeaux, Aquitaine, France
Matthias Stuber, PhD
Professor
University Hospital (CHUV) and University of Lausanne (UNIL)
Lausanne, Switzerland
Hubert Cochet, MD, PhD
PROF/PhD
Centre Hospitalier Universitaire de Bordeaux
Pessac, Aquitaine, France
Bright-blood late gadolinium-enhanced (LGE) imaging is the current clinical gold standard to assess myocardial scar1. However, poor contrast at the blood-scar interface makes scar detection and quantification challenging. We introduce a framework for fully automated scar detection and quantification combining novel joint bright- and black-blood LGE imaging (SPOT) with artificial intelligence (AI)-powered analysis.
Methods:
Acquisition: 300 patients (83 females, 55±15 years) with prior myocardial infarction underwent CMR at 1.5T (Siemens, Aera). Short-axis 2D whole-heart reference PSIR2 and SPOT3 images were collected under breath-hold in random order 15min after injection of gadoteric acid. The SPOT sequence jointly collects bright-blood (wall visualization) and black-blood (scar visualization) images by combining inversion-recovery with T1-rho pulses (Fig1).
Processing: Two radiologists provided ground truth left ventricular (LV) wall and scar contours on both PSIR and SPOT images using CVI42 (Circle, Calgary). A 2D U-net4 model was trained on a subset of 288 patients (~4300 images) to automatically extract wall contours from bright-blood SPOT images. Wall contours were then propagated to black-blood SPOT images and scar was extracted as voxels with signal intensity >12 standard deviation above the mean intensity in a region-of-interest automatically defined in the LV cavity. Total and regional infarct mass, volume (% of the wall), and transmurality were automatically measured.
Experiments: In 12 patients (4 females, age 61±10 years) unseen during training, Bland-Altman analyses and correlation coefficients were used to assess inter-reader and inter-method agreements. Confusion matrices were used to study the accuracy of the framework to detect scar and to grade its transmurality across AHA segments.
Results:
SPOT datasets were automatically processed in ~10s (vs. 17±7min for manual processing). As compared to PSIR, SPOT images showed higher inter-reader agreement in manual LV wall segmentation (R2=0.92, ICC=0.83 vs. R2=0.79, ICC=0.58) (Fig2A). The trained U-net allowed for accurate automated LV wall segmentation on bright-blood SPOT images, with a non-significant bias between manual and automated LV mass values (-2.5g, 95%CI: -20 to +15g, R2=0.92, P=0.35) (Fig2B). Applying these LV wall contours to black-blood images resulted in a fully automated quantification of scar mass, with excellent correlation to manual measurements (R2=0.93, bias: -0.39g, 95%CI: -9.5 to +8.7g, P< 0.01) (Fig2B). This automated scar quantification compared favourably to gold standard manual PSIR quantifications (Fig2C). On a segmental basis, and as compared to labour intensive PSIR expert segmentation, the AI-powered SPOT approach allowed for automated mapping of scar presence (89% accuracy) and transmurality (75% accuracy) across AHA segments (Fig2D & 3).
Conclusion:
The AI-powered SPOT imaging framework allows for fast, robust, and fully automated detection and quantification of post-infarction scar.