Cardiac Segmentation With Strong Anatomical Guarantees.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Cardiology

Background:

• Convolutional neural networks (CNNs) excel in medical image segmentation but can produce anatomically inaccurate results.
• Existing methods may struggle with anatomical plausibility, even with shape priors.

Purpose of the Study:

• To develop a framework for cardiac image segmentation that guarantees anatomical correctness while maintaining inter-expert variability.
• To ensure segmentation results are both precise and anatomically plausible across different imaging modalities.

Main Methods:

• Utilizing a well-trained CNN for initial cardiac image segmentation.
• Employing a constrained variational autoencoder (cVAE) to identify and correct anatomically implausible segmentations.
• Warping inaccurate shapes to the nearest valid cardiac shape within a learned latent space.

Main Results:

• The framework successfully produced cardiac segmentation maps that are anatomically plausible.
• Results remained within inter-expert variability across diverse modalities like MRI and ultrasound.
• Eliminated the need for shape priors to ensure anatomical correctness.

Conclusions:

• The proposed method enhances CNN-based cardiac image segmentation by ensuring anatomical plausibility.
• This approach offers a robust solution for accurate and reliable cardiac image analysis.
• The framework is adaptable to different cardiac imaging modalities.