Automated Cardiac MR Ischemic Scar Assessment

A deep learning-based system for the accurate, time-efficient segmentation and clinical reporting of left ventricular myocardial infarction scar.

Assessing myocardial infarction scar (MIS) through Cardiovascular Magnetic Resonance (CMR) provides critical prognostic information and guides clinical management. However, manual MIS segmentation is highly time-consuming and prone to reproducibility issues.

To address this bottleneck, we developed a deep learning-based computational workflow that automates the segmentation of the left ventricle (LV) and MIS. For the first time, this model was trained and deployed on state-of-the-art Dark Blood Late Gadolinium Enhancement (DB-LGE) images. The system not only performs highly accurate segmentations but also automatically generates a quantitative clinical report calculating MIS extent and transmurality.

Cardiology U-Net CNN Clinical Automation

The Clinical Challenge

The presence, location, and extent of MIS correlate strongly with arrhythmia risk and the likelihood of functional recovery following revascularization. While CMR LGE imaging is the gold standard for detecting these scars, clinical adoption is limited by the lack of an optimal, standardized quantification method.

  • Manual contouring: Lacks reproducibility and is operator-dependent.
  • Semi-automated techniques: (e.g., n-Standard Deviation thresholds) improve reproducibility but still require significant manual intervention and take an average of 7±3 minutes per patient.
  • Previous AI attempts: Mostly relied on conventional Bright Blood (BB) LGE images, where low contrast between the blood pool and the scar complicates the identification of thin, subendocardial scars.

A Novel Deep Learning Approach

Our solution utilizes state-of-the-art Dark Blood (DB) LGE sequences, which suppress the signal from the blood pool to maximize contrast and sensitivity for detecting subendocardial MIS.

We designed a sequential pipeline utilizing two Convolutional Neural Networks (CNNs) based on the U-Net architecture:

  1. Left Ventricle Segmentation: The first model accurately delineates the LV endocardial and epicardial borders.
  2. MIS Segmentation: The second model isolates the specific regions of myocardial infarction scar within the identified LV myocardium.

Performance & Clinical Impact

The system was trained and validated on a robust, bi-centric dataset comprising 1,355 DB-LGE short-axis images from 144 patients.

  • High Accuracy: The models achieved strong performance metrics.
  • Time Efficiency: Compared to the semi-automated 4-Standard Deviation technique, our deep learning system was five times faster (completing the analysis in under 1 minute) and required minimal user interaction.
  • Automated Reporting: The workflow automatically computes clinically vital metrics—MIS extent (relative infarcted mass) and transmurality—and formats them into a ready-to-use clinical report.

Research Highlights

Dark Blood LGE

Pioneering the use of deep learning on DB-LGE sequences to maximize contrast and detect thin subendocardial scars.

5x Faster Workflow

Reduces image processing and segmentation time from ~7 minutes to under 1 minute per patient.

Automated Reporting

Calculates transmurality and relative infarcted mass, directly generating a quantitative clinical report.

Request Information

Core Publication

2023

  1. An accurate and time-efficient deep learning-based system for automated segmentation and reporting of cardiac magnetic resonance-detected ischemic scar
    Daniele M Papetti, Kirsten Van Abeelen, Rhodri Davies, and 8 more authors
    Computer methods and programs in biomedicine, 2023
    DOI