Direct Multitype Cardiac Indices Estimation via Joint Representation and Regression Learning.

Area of Science:

• Cardiovascular Imaging
• Medical Image Analysis
• Deep Learning

Background:

• Accurate cardiac indices estimation is crucial for diagnosing cardiac diseases.
• Current methods struggle with cardiac structure variability and temporal dynamics in MRI, leading to unreliable estimations.
• Existing approaches often use vulnerable feature engineering and incompatible regression models.

Purpose of the Study:

• To develop a semi-automated, highly accurate method for estimating multitype cardiac indices from cardiac MR sequences.
• To address the limitations of existing methods in feature representation and model compatibility.
• To enhance the clinical utility of cardiac MRI for disease assessment.

Main Methods:

• Proposed a semi-automated method utilizing an integrated deep neural network, Indices-Net.
• Indices-Net employs a deep convolution autoencoder for image representation and a multiple output convolution neural network for regression.
• The network jointly learns representation and regression models after manual landmark labeling for Region of Interest (ROI) cropping.

Main Results:

• Indices-Net achieved low estimation errors for LV wall thicknesses (1.44 ± 0.71 mm) and cavity/myocardium areas (204 ± 133 mm²).
• Demonstrated significant error reductions compared to segmentation (55.1%, 17.4%) and two-phase volume-only methods (12.7%, 14.6%).
• Validated on 145 subjects using five-fold cross-validation.

Conclusions:

• The proposed Indices-Net method offers accurate and reliable estimation of multitype cardiac indices.
• Joint learning enhances image representation expressiveness and model compatibility for improved cardiac index prediction.
• This method shows great potential for clinical cardiac function assessment and diagnosis.