Automated segmentation of biventricular contours in tissue phase mapping using deep learning.

Area of Science:

• Cardiovascular Imaging
• Artificial Intelligence in Medicine
• Medical Image Analysis

Background:

• Tissue phase mapping (TPM) is an MRI technique for quantifying biventricular myocardial velocities.
• Clinical application of TPM is hindered by the time-consuming manual segmentation of cardiac contours.
• Automated segmentation methods are needed to improve the efficiency and accessibility of TPM analysis.

Purpose of the Study:

• To develop and validate a deep learning (DL) network for automated segmentation of TPM images.
• To assess the accuracy of DL-based segmentation compared to manual segmentation.
• To evaluate the impact of automated segmentation on myocardial velocity quantification.

Main Methods:

• A multi-channel 3D dense U-Net architecture was implemented for segmentation.
• The network was trained on magnitude and phase TPM images, using combined loss functions (cross-entropy, Dice, Hausdorff distance).
• The DL model was trained and tested on 150 TPM scans from heart transplant patients, with manual segmentation masks as ground truth.

Main Results:

• The DL network achieved a median Dice score of 0.85 for the left ventricle and 0.64 for the right ventricle.
• Segmentation time was reduced by approximately 3,600-fold, from 2 hours manually to 1.9 seconds with DL.
• Peak radial (Vr) and longitudinal (Vz) velocities derived from DL segmentation showed strong correlation (R ≥ 0.88) and good agreement with manual analysis.

Conclusions:

• The proposed 3D dense U-Net enables highly accurate and significantly faster automated segmentation of TPM images.
• This DL approach minimizes the loss of accuracy in myocardial velocity quantification compared to manual segmentation.
• The automated method has the potential to overcome clinical limitations and facilitate wider adoption of TPM in cardiac assessment.