Multi-scale deep networks and regression forests for direct bi-ventricular volume estimation.

Area of Science:

• Medical Imaging
• Artificial Intelligence in Medicine
• Cardiovascular Imaging

Background:

• Direct estimation of cardiac ventricular volumes is crucial for cardiac function analysis.
• Current methods often require extensive user input or rely on flawed assumptions.
• There is a need for automated, reliable methods for ventricular volume estimation.

Purpose of the Study:

• To present a general, fully learning-based framework for direct bi-ventricular volume estimation.
• To eliminate the need for user inputs and problematic assumptions in volume estimation.
• To improve the accuracy and efficiency of cardiac function analysis.

Main Methods:

• A two-stage learning framework was developed: unsupervised cardiac image representation learning using multi-scale deep networks and direct bi-ventricular volume estimation using random forests.
• The framework leverages both generative and discriminant learning approaches.
• A leave-one-subject-out cross-validation was employed.

Main Results:

• The proposed method achieved high correlations (around 0.92) with ground truth for both left and right ventricles.
• It significantly outperformed existing direct estimation methods on a dataset of 100 subjects (healthy and diseased).
• The method demonstrated robustness and accuracy in a larger, more diverse cohort.

Conclusions:

• The fully learning-based framework offers a practical and accurate solution for direct bi-ventricular volume estimation in clinical settings.
• This approach removes reliance on user input and assumptions, enhancing reliability.
• The framework is adaptable for estimating volumes of other organs.