A distance map regularized CNN for cardiac cine MR image segmentation.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Cardiovascular Research

Background:

• Cardiac image segmentation is vital for personalized heart models and performance quantification.
• Accurate segmentation of cardiac structures from cine MR images is challenging due to anatomical variability and diverse imaging protocols.

Purpose of the Study:

• To propose a multi-task learning (MTL)-based regularization of a convolutional neural network (CNN) for accurate cardiac structure segmentation from cine MR images.
• To improve the accuracy and robustness of automatic segmentation of the left ventricle (LV), right ventricle (RV), and myocardium.

Main Methods:

• A CNN was trained for semantic segmentation and pixel-wise distance map regression simultaneously.
• Task losses were weighted by inverse uncertainties, allowing dynamic adjustment during training.
• A decoder network was added as a regularizer to learn global features, removed post-training.

Main Results:

• The proposed method significantly improved both binary and multi-class segmentation performance compared to state-of-the-art CNNs.
• Achieved average Dice coefficients of 0.84 ± 0.03 and 0.91 ± 0.04 on public datasets.
• Demonstrated a 42% improvement in myocardium Dice coefficient (0.56 to 0.80) in cross-dataset segmentation tasks.

Conclusions:

• The developed method achieves accurate cardiac structure segmentation from cine MR images, outperforming existing methods.
• Cardiac indices derived from the segmentation show strong correlation with ground truth, supporting clinical utility.
• The method shows potential as a non-invasive diagnostic tool and for patient-specific cardiac modeling in therapy planning.