Myocardium Detection by Deep SSAE Feature and Within-Class Neighborhood Preserved Support Vector Classifier and...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Cardiovascular Research

Background:

• Left ventricle myocardium detection is crucial for cardiac analysis but challenging due to anatomical variability.
• Accurate segmentation and registration of cardiac structures are essential for quantitative assessments.

Purpose of the Study:

• To develop and validate a hybrid deep learning model for robust automatic detection of left ventricle myocardium in cardiac MRI.
• To enhance the accuracy and generalization of myocardium localization for improved cardiac image processing.

Main Methods:

• A hybrid model combining region proposal (supervoxel over-segmentation, hierarchical clustering) and deep feature classification/regression (deep stacked sparse autoencoder, C-SVC, ε-SVR).
• Incorporation of hard negative sample mining to refine the deep stacked sparse autoencoder and classifier for improved stability and generalization.
• Extensive evaluation on a public cardiac dataset, comparing the integrated components and the overall model against state-of-the-art methods.

Main Results:

• The proposed hybrid model demonstrated robust performance in myocardium localization.
• Experimental results verified the effectiveness of the integrated components and showed superior performance compared to existing methods.
• The model achieved high accuracy in typical metrics for cardiac image analysis.

Conclusions:

• The developed hybrid model offers an effective solution for automatic left ventricle myocardium detection in cardiac MRI.
• This approach provides significant benefits for downstream cardiac image processing tasks, including region-of-interest cropping and left ventricle volume measurement.
• The study highlights the potential of deep learning for advancing quantitative cardiovascular imaging analysis.