Anatomically-informed deep learning on contrast-enhanced cardiac MRI for scar segmentation and clinical feature extraction
Inventors
Trayanova, Natalia A. • ABRAMSON, Haley Gilbert • Popescu, Dan • MAGGIONI, Mauro • WU, Katherine C.
Interested in licensing this patent?
MTEC can help explore whether this patent might be available for licensing for your application.
Assignees
MemberJohns Hopkins UniversityJohns Hopkins UniversityFounded in 1876, Johns Hopkins University is recognized as the first research university in the United States. It advances interdisciplinary education, high-impact research, and global outreach, supporting knowledge translation, technological innovation, and community partnerships. The university fosters academic excellence, innovation incubation, outreach, and inclusion across multiple campuses in Baltimore, integrating into the city's social, economic, and cultural life.
Founded in 1876, Johns Hopkins University is recognized as the first research university in the United States. It advances interdisciplinary education, high-impact research, and global outreach, supporting knowledge translation, technological innovation, and community partnerships. The university fosters academic excellence, innovation incubation, outreach, and inclusion across multiple campuses in Baltimore, integrating into the city's social, economic, and cultural life.
Publication Number
US-12670598-B2
Publication Date
2026-06-30
Expiration Date
Abstract
Fully automated computer-implemented deep learning techniques of contrast-enhanced cardiac MRI segmentation are provided. The techniques may include providing cardiac MRI data to a first computer-implemented deep learning network trained in order to identify a left ventricle region of interest to generate left ventricle region-of-interest-identified cardiac MRI data. The techniques may also include providing the left ventricle region-of-interest-identified cardiac MRI data to a second computer-implemented deep learning network trained in order to identify myocardium to generate myocardium-identified cardiac MRI data. The techniques may further include providing the myocardium-identified cardiac MRI data to at least one third computer-implemented deep learning network trained to conform data to geometrical anatomical constraints in order to generate anatomical-conforming myocardium-identified cardiac MRI data. The techniques may further include outputting the anatomical-conforming myocardium-identified cardiac MRI data.
Core Innovation
The invention provides a fully automated computer-implemented deep learning method for contrast-enhanced cardiac MRI segmentation. Cardiac MRI data is provided to a first computer-implemented deep learning network trained to identify a left ventricle region of interest, producing left ventricle region-of-interest-identified cardiac MRI data. The first network comprises a convolutional neural network with residuals.
The left ventricle region-of-interest-identified cardiac MRI data is provided to a second computer-implemented deep learning network trained to identify myocardium, producing myocardium-identified cardiac MRI data. The second network comprises a convolutional neural network with residuals. The myocardium-identified data is then provided to at least one third computer-implemented deep learning network trained to conform data to geometrical anatomical constraints to ensure anatomical accuracy.
The at least one third computer-implemented deep learning network comprises a convolutional autoencoder coupled to a Gaussian mixture model. The third network is trained to conform data to geometrical anatomical constraints by autoencoding the myocardium-identified cardiac MRI data to generate a latent vector space and statistically modeling the latent vector space. The latent vector space is populated with anatomically-correct samples and allows for nearest-neighbor identification of the anatomical-conforming myocardium-identified cardiac MRI data, whereby the anatomical-conforming myocardium-identified cardiac MRI data is anatomically correct, and no manual human intervention is required.
Claims Coverage
The partial claims include two independent claims (clm-00001 and clm-00006). Each independent claim includes three main inventive features: residual convolutional neural networks for left ventricle region-of-interest identification and myocardium identification, and an anatomical-conforming stage that uses a convolutional autoencoder coupled to a Gaussian mixture model to perform latent-space nearest-neighbor identification under geometrical anatomical constraints, with output providing anatomically correct myocardium-identified cardiac MRI data without manual human intervention.
Fully automated contrast-enhanced cardiac MRI segmentation with residual convolutional neural networks
Providing cardiac MRI data to a first computer-implemented deep learning network comprising a convolutional neural network with residuals to identify a left ventricle region of interest and produce left ventricle region-of-interest-identified cardiac MRI data, and providing the left ventricle region-of-interest-identified cardiac MRI data to a second computer-implemented deep learning network comprising a convolutional neural network with residuals to identify myocardium and produce myocardium-identified cardiac MRI data.
Anatomical conformity using a convolutional autoencoder coupled to a Gaussian mixture model
Conforming the myocardium-identified cardiac MRI data to geometrical anatomical constraints by using at least one third computer-implemented deep learning network comprising a convolutional autoencoder coupled to a Gaussian mixture model, wherein the network is trained by autoencoding to generate a latent vector space and statistically modeling the latent vector space.
Latent-space nearest-neighbor identification for anatomically correct output without manual intervention
Populating the latent vector space with anatomically-correct samples and allowing nearest-neighbor identification of the anatomical-conforming myocardium-identified cardiac MRI data, whereby the anatomical-conforming cardiac MRI data is anatomically correct, and outputting the anatomical-conforming myocardium-identified cardiac MRI data wherein no manual human intervention is required.
Across both independent claims, the coverage centers on a fully automated deep learning pipeline that first identifies a left ventricle region of interest and then identifies myocardium using convolutional neural networks with residuals. The method and system then ensure anatomical accuracy by conforming myocardium-identified cardiac MRI data to geometrical anatomical constraints using a convolutional autoencoder coupled to a Gaussian mixture model, with nearest-neighbor identification in a latent vector space populated with anatomically-correct samples, producing anatomically correct output without manual human intervention.
Stated Advantages
Fully automated segmentation wherein no manual human intervention is required.
Documented Applications
Contrast-enhanced cardiac MRI segmentation, including left ventricle region-of-interest identification, myocardium identification, and anatomical-conforming myocardium-identified cardiac MRI data output.
Scar segmentation data included as an anatomical-conforming output in dependent claims.
Reducing background in ventricle region-of-interest-identified cardiac MRI data as an additional processing refinement in dependent claims.
Delineating myocardium by delineating endocardium and epicardium as an additional refinement in dependent claims.
Displaying the anatomical-conforming myocardium-identified cardiac MRI data on a computer monitor as an additional output refinement in dependent claims.
Interested in licensing this patent?