Segmentation of Left Ventricle From 3D Cardiac MR Image Sequences Using A Subject-Specific Dynamical Model

Area of Science:

• Medical Imaging
• Computational Anatomy
• Biomedical Engineering

Background:

• Cardiac image segmentation is crucial for diagnosing heart conditions.
• Existing statistical models often fail to capture both inter-subject shape variability and temporal cardiac motion dynamics.
• Static models (SM) ignore temporal coherence, while generic dynamical models (GDM) overlook subject-specific variations.

Purpose of the Study:

• To introduce a subject-specific dynamical model (SSDM) for accurate left ventricle segmentation in cardiac imaging.
• To simultaneously address inter-subject variability and intra-subject temporal dynamics.
• To improve the consistency and accuracy of cardiac segmentation over time.

Main Methods:

• Developed a subject-specific dynamical model (SSDM) integrating inter-subject variability and temporal dynamics.
• Employed a recursive Bayesian framework to combine current frame intensity information with past frame predictions.
• Validated the SSDM approach using a "Leave-one-out" cross-validation strategy on 32 cardiac image sequences.

Main Results:

• The SSDM demonstrated superior performance in segmenting the left ventricle compared to static models (SM) and generic dynamical models (GDM).
• Quantitative analysis confirmed better global and local consistencies of SSDM segmentations against manual segmentations.
• The model effectively identified and adapted to specific motion patterns within individual cardiac sequences.

Conclusions:

• The subject-specific dynamical model (SSDM) offers a significant advancement in statistical model-based cardiac image segmentation.
• SSDM effectively handles both inter-subject shape variations and intra-subject cardiac motion.
• This approach leads to more accurate and consistent segmentation of the left ventricle in cardiac sequences.