Atrial scar quantification via multi-scale CNN in the graph-cuts framework.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Cardiology

Background:

• Late gadolinium enhancement magnetic resonance imaging (LGE MRI) is crucial for assessing myocardial scar in atrial fibrillation (AF).
• Automating scar quantification in LGE MRI is challenging due to low image quality and complex atrial anatomy.
• Accurate scar assessment is vital for the diagnosis and prognosis of AF.

Purpose of the Study:

• To develop and validate a fully automated method for quantifying atrial scars using LGE MRI.
• To improve the accuracy and efficiency of scar segmentation in the left atrium (LA).
• To compare the performance of the proposed automated method against conventional manual approaches.

Main Methods:

• A novel automated method employing the graph-cuts framework was developed.
• Graph potentials were learned on a left atrium (LA) surface mesh using a multi-scale convolutional neural network (MS-CNN).
• The method was validated using 58 LGE MRI images with manual delineations, focusing on balancing graph cut weights.

Main Results:

• The proposed MS-CNN integrated with graph-cuts significantly improved LA scar segmentation accuracy.
• The automated method achieved a mean accuracy of 0.856 ± 0.033 and a mean Dice score of 0.702 ± 0.071.
• The automated method demonstrated significantly superior performance (p < 0.01) compared to conventional manual delineation methods.

Conclusions:

• The fully automated graph-cuts and MS-CNN based method offers a promising solution for accurate LA scar quantification in LGE MRI.
• This technique can enhance the diagnosis and prognosis of atrial fibrillation.
• The method's ability to incorporate both local and global image textures contributes to its improved segmentation accuracy.