Register now After registration you will be able to apply for this opportunity online.
This opportunity is not published. No applications will be accepted.
Machine Learning for Ventricular Scar Segmentation on LGE CMR Images
The aim of the project is to develop a machine learning model for ventricular scar segmentation on Late Gadolinium Enhanced Cardiac Magnetic Resonance (LGE CMR) images.
Keywords: machine learning, ML, deep learning, neural networks, MRI, image processing, AI, data analysis
Ventricular scar typically forms when a section of the myocardial wall is deprived of blood, most commonly caused by myocardial infarction. LGE CMR is the gold standard for scar imaging. It results in images with a contrast between normal myocardium (dark/no signal) and the scar tissue (bright). Currently, conventional threshold-based methods (e.g., 2-SD and FWHM) for ventricular scar segmentation are widely used by clinicians. However, they are unable to handle common variabilities of the scars, such as their morphology and brightness distribution. Hence, deep learning methods are emerging, which can handle variability well and improve scar segmentation accuracy.
Ventricular scar typically forms when a section of the myocardial wall is deprived of blood, most commonly caused by myocardial infarction. LGE CMR is the gold standard for scar imaging. It results in images with a contrast between normal myocardium (dark/no signal) and the scar tissue (bright). Currently, conventional threshold-based methods (e.g., 2-SD and FWHM) for ventricular scar segmentation are widely used by clinicians. However, they are unable to handle common variabilities of the scars, such as their morphology and brightness distribution. Hence, deep learning methods are emerging, which can handle variability well and improve scar segmentation accuracy.
Currently, we are able to segment scar tissue from short-axis views. However, inter-slice information should be leveraged to improve the segmentation accuracy and obtain spatially continuous 3D scar segmentations. The aim of this work is to develop a machine learning model that incorporates both intra-slice and inter-slice information for an accurate 3D ventricular scar segmentation. The current 2D segmentation model should be investigated given the available data, and limitations from only leveraging the information given in the short-axis view should be highlighted. Accordingly, a 3D CNN model or a model using building blocks from recursive neural networks will be implemented to include inter-slice information. The results of the model should be evaluated on available labeled LGE CMR images.
Currently, we are able to segment scar tissue from short-axis views. However, inter-slice information should be leveraged to improve the segmentation accuracy and obtain spatially continuous 3D scar segmentations. The aim of this work is to develop a machine learning model that incorporates both intra-slice and inter-slice information for an accurate 3D ventricular scar segmentation. The current 2D segmentation model should be investigated given the available data, and limitations from only leveraging the information given in the short-axis view should be highlighted. Accordingly, a 3D CNN model or a model using building blocks from recursive neural networks will be implemented to include inter-slice information. The results of the model should be evaluated on available labeled LGE CMR images.