CNN-based fully automatic mitral valve extraction using CT images and existence probability maps.

Area of Science:

• Medical Imaging
• Cardiovascular Surgery
• Artificial Intelligence in Medicine

Background:

• Accurate mitral valve shape extraction from tomographic images is crucial for planning surgical and interventional treatments.
• Manual extraction is time-consuming, and existing automated methods lack sufficient accuracy.
• Mitral valve segmentation is challenging due to variations across the cardiac cycle and patient anatomy.

Purpose of the Study:

• To develop a fully automated method for extracting mitral valve shape from computed tomography (CT) images across all cardiac phases.
• To improve the accuracy of mitral valve shape extraction compared to existing methods.
• To provide a tool for enhanced planning of mitral valve interventions.

Main Methods:

• A novel method utilizing DenseNet and U-Net was developed for mitral valve shape extraction.
• The approach incorporates original CT images and U-Net-inferred probability maps of the mitral valve area.
• The method was trained and validated on 1585 CT images from 204 patients using 10-fold cross-validation.

Main Results:

• The proposed automated method achieved a mean shape extraction error of 0.88 mm.
• This represents a significant improvement of 0.32 mm compared to a method without probability maps.
• The use of existence probability maps enhanced the accuracy of mitral valve segmentation.

Conclusions:

• A novel, fully automated method for mitral valve shape extraction from 4D CT images has been presented.
• The integration of existence probability maps demonstrably improves the accuracy of mitral valve shape extraction.
• This advancement holds potential for more precise pre-procedural planning in cardiac interventions.