AI-powered SPOT imaging for enhanced myocardial scar detection and quantification

Abstract

Cardiovascular disease is the leading global cause of death, underscoring the need for accurate assessment of myocardial injury. The current gold standard, bright-blood late gadolinium enhanced MRI, suffers from poor contrast at the blood-scar interface, reducing sensitivity for subendocardial scar detection and limiting reproducibility. Moreover, reliance on expert manual analysis makes interpretation labor-intensive and variable. Here, we present SPOT, a multi-spectral bright- and black-blood imaging sequence that provides unprecedented scar-to-blood contrast and clear anatomical detail. Integrated with an artificial intelligence (AI) framework for automated image analysis, SPOT enables rapid, fully automated, and operator-independent quantification of myocardial injury. Validated in simulations, animal infarct models, and patients with heart disease, this combined imaging-AI platform delivers accurate detection and quantification in a single acquisition. This innovation presents significant opportunities for earlier diagnosis and enhanced therapeutic management of ischemic heart disease, with potential applications in a wide spectrum of other clinical settings.

Fig. 1. Overview of the SPOT image acquisition, segmentation, and scar analysis processes.

a SPOT collects 2D co-registered bright-blood and black-blood late gadolinium enhancement images in a single sequence. A transformer-based model is employed to fully automatically segment the left ventricular wall on bright-blood SPOT images, and a U-net is used to detect the two right ventricular insertion points. The segmented contours and identified landmarks are then propagated onto the black-blood images, facilitating detailed myocardial scar analysis, including the assessment of scar size and transmurality. The final reporting includes key metrics such as LV wall volume, scar volume, scar burden, and affected AHA segments. b Comparative short-axis SPOT and PSIR images collected in a patient with subendocardial septal scar, along with the signal evolution across the interventricular septum. The corresponding signal intensity profiles across the interventricular septum are illustrated, highlighting the superior blood-scar contrast achieved with SPOT black-blood imaging compared to PSIR. The schematic representations below each image depict the differential signal profiles and the relative effectiveness of each imaging sequence in distinguishing between blood pool, scar tissue, and healthy myocardium.

Fig. 2. Extended phase graph simulations and phantom validation of SPOT imaging.

a Simulations explore the interaction between inversion time and T1-rho pulse duration on signal dynamic and contrast enhancement of myocardium, blood, and scar tissues. The top panel illustrates simulations for odd heartbeats (black-blood image), revealing an optimal nulling of both blood and healthy myocardium signals achieved with a T1-rho duration of 27 msec. The bottom panel displays simulations for even heartbeats (bright-blood image), showcasing a favorable myocardium-blood contrast achieved with a T1-rho duration of 50 msec. The panels to the right show representative black-blood and bright-blood images illustrating the effects of varying T1-rho durations on myocardial scar visibility. b Results from the T1MES Phantom experiment comparing conventional bright-blood PSIR imaging with the proposed bright- and black-blood SPOT images. Signal intensities were measured in vials with T1/T2 values corresponding to post-contrast blood (green circle), viable myocardium (red circle), and scar tissues (yellow circle). Conventional PSIR imaging effectively suppresses the signal from the viable myocardium compartment but fails to suppress the blood signal, resulting in limited scar contrast. In contrast, the black-blood SPOT image demonstrates clear signal attenuation in both blood and viable myocardium compartments, yielding superior scar-to-myocardium and scar-to-blood contrasts. Signal intensity measurements further corroborate these observations, with distinct differences observed across the three imaging modalities shown in the lower panel.

Fig. 3. Comparative analysis of SPOT imaging with histology and gross pathology in animal models.

This figure presents a comprehensive comparison of phase-sensitive inversion recovery (PSIR), SPOT imaging, Masson trichome histology, and gross pathology in two distinct ovine models of myocardial injury. Top row: imaging results from a sheep model subjected to multiple radiofrequency ablations, mimicking microinfarctions. The PSIR images show limited contrast in delineating the ablation sites, while the SPOT black-blood images enhance the visibility of the scar tissue. The bright-blood fusion images further improve the localization of the scar, with clear demarcation observed in the corresponding histology and gross pathology sections. This experiment was performed in one sheep. Scale bars: histology (left) 5000 μm, histology (right) 1000 μm, gross pathology 5 mm. Bottom row: imaging results from a sheep model with a large myocardial infarction induced by ischemia-reperfusion via balloon occlusion. The PSIR images again show suboptimal contrast for scar detection. In contrast, SPOT black-blood images provide superior scar delineation, while the bright-blood fusion images accurately localize the infarcted regions. These imaging results are validated by the corresponding histology and gross pathology, which reveal extensive fibrosis in the infarcted myocardium. This experiment was performed in one sheep. Representative data are shown; experiments were conducted in two sheep in total. Scale bars: histology (left) 500–5000 μm (depending on panel), histology (right) 500–5000 μm (depending on panel), gross pathology 5 mm. Arrows highlight myocardial injuries. MI myocardial infarction, RF radiofrequency.

Fig. 4. Evaluation of image quality, segmentation accuracy, signal intensities, and DICE scores.

a Illustrative examples of artifacts observed in SPOT images. These artifacts primarily manifested as blood flow artifacts, predominantly affecting the basal slices. Additionally, susceptibility artifacts linked with balanced steady-state free-precession readouts were identified. Residual motion artifacts were also observed, particularly in patients experiencing challenges in maintaining consistent breath-holds. Artifacts are highlighted by arrows for clarity. b Automated left ventricular wall segmentation results and detected right ventricular insertion points on bright-blood SPOT images in six patients from the testing set. The segmentation outcomes demonstrate the ability of the automated pipeline to accurately delineate the left ventricle and reliably identify right ventricular insertion points (A-RVI and P-RVI), as depicted by the overlay of segmentation contours and landmarks. c Accuracy assessment of automatically detected right ventricular insertion points, reported as median values with interquartile ranges (25th–75th percentile). The reference point to compute the insertion points angle was the left ventricular center of mass. Two-sided Wilcoxon signed rank test. Confidence score, P = 4.6 × 10⁻⁶; Euclidian distance, P = 0.295; insertion points angle P = 0.924. A-RVI anterior right ventricular insertion point, P-RVI posterior right ventricular insertion point. d Comparative analysis of mean signal intensities obtained in scar, myocardium, and blood regions across the testing set cohort. Signal intensities were evaluated in reference PSIR images and proposed bright- and black-blood SPOT images (N = 50). The two-sided Wilcoxon signed rank test was employed to compare signal intensities between tissue types. Reference PSIR: scar versus blood: P = 0.171; scar versus myocardium: P = 1.6 × 10⁻⁸; blood versus myocardium: P = 1.6 × 10⁻⁸; SPOT black-blood: scar versus blood: P = 2.4 × 10⁻⁸; scar versus myocardium: P = 2.4 × 10⁻⁸; blood versus myocardium: P = 0.125; SPOT bright-blood: scar versus blood: P = 5.9 × 10⁻⁸; scar versus myocardium: P = 6.9 × 10⁻³; blood versus myocardium: P = 2.4 × 10⁻⁸. e Comparison of myocardial scar segmentation techniques for black-blood SPOT imaging using the DICE index, n = 44 independent patients. (Left) Comparison of segmentation performance between the full width at half maximum (FWHM) method, various n-SD approaches (n = 1–4) and the region growing technique against manual segmentation for infarct size quantification on black-blood SPOT images. For the n-SD segmentation method (n = 1–4), the reference signal was manually extracted from a region-of-interest delineated in the remote (non-infarcted) myocardium. The starting seed point for the seeded region growing algorithm was automatically determined by the left ventricular center of mass, eliminating the need for user interaction. (Right) Multiple region growing distances ε were evaluated to optimize the accuracy of infarct delineation. Box plots display the 25th and 75th percentiles (bounds of the box), with the center line indicating the median (50th percentile). Whiskers extend to the minimum and maximum values, and individual points represent outliers. Source data are provided as a Source data file.

Fig. 5. Comparative scar segmentation results and whole ventricle visualization in patients with myocardial infarction.

a Comparison of full width at half maximum (FWHM), n-SD (n ranging from 1 to 4) and the region growing technique against manual segmentation for infarct size quantification. The overlay of segmentation contours on the SPOT images illustrates the variability and accuracy of each method in delineating the infarcted regions. b Comparison of automated segmentation using a seeded region growing algorithm against gold standard manual segmentation across four patients of varying ages and genders. The top row shows black-blood SPOT images, followed by automated segmentation results in the middle row, and manual segmentation outcomes in the bottom row. c Whole ventricle visualization of myocardial scar in a 43-year-old male patient with myocardial infarction originating from the left anterior descending artery. The figure compares SPOT and PSIR images, providing a slice-by-slice assessment of scar extent as a percentage of left ventricular mass. The corresponding bullseye plots illustrate the infarct size and transmurality across different segments of the left ventricle.

Fig. 6. Illustrative SPOT images collected in patients and distribution of LGE segments.

a Illustrative examples of SPOT images collected from three patients with myocardial infarction, demonstrating the challenges in accurately delineating subendocardial and scar contours using reference phase-sensitive inversion recovery (PSIR) imaging. Yellow arrows indicate scar areas that are difficult to delineate on PSIR images. b Illustrative PSIR and SPOT images from the testing set collected on six patients with large myocardial scars. Yellow arrows highlight scar areas that are difficult to delineate on PSIR images, often due to poor contrast or ambiguous boundaries. SPOT imaging clearly delineates these scar regions. c PSIR and SPOT images from six patients with small, localized myocardial scars. Yellow arrows indicate scar locations, which are difficult to detect on PSIR images due to their subtlety and proximity to the blood pool. d Distribution of LGE segments across the entire testing set, represented in American Heart Association (AHA) bullseye plots. The regional distribution of LGE segments is shown for reference PSIR imaging, alongside comparisons with manual and fully automated SPOT analyses. The plots illustrate the superior detection and localization of LGE segments using SPOT imaging, both with manual and automated assessments.