Joint Segmentation and Uncertainty Estimation of Ventricular Structures from Cardiac MRI using a Bayesian...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Cardiology

Background:

• Convolutional Neural Networks (CNNs) show promise for cardiac MRI segmentation but lack clinical reliability.
• Quantifying segmentation uncertainty is crucial for identifying unreliable outputs, especially with real-world data variations.
• Accurate uncertainty estimates are vital for determining the trustworthiness of AI model predictions in clinical practice.

Purpose of the Study:

• To develop a Bayesian deep-learning framework for reliable cardiac structure segmentation from MRI.
• To quantify predictive uncertainty in cardiac image segmentation to enhance clinical trust.
• To evaluate the correlation between segmentation uncertainty and errors using a novel deep-learning model.

Main Methods:

• A Bayesian version of the CondenseUNet framework was employed, incorporating a learned group structure and a regularized weight-pruner.
• The model was designed to reduce computational costs in volumetric image segmentation while enabling uncertainty quantification.
• The framework was trained and validated on the Automated Cardiac Diagnosis Challenge (ACDC) dataset.

Main Results:

• The proposed Bayesian CondenseUNet effectively segmented cardiac structures (left ventricle, right ventricle, myocardium) in cine cardiac MRI.
• The study demonstrated the model's capability to correlate segmentation uncertainty with actual segmentation errors.
• The framework provided well-calibrated uncertainty estimates, indicating the reliability of segmentation outputs.

Conclusions:

• The Bayesian CondenseUNet offers a robust solution for reliable cardiac MRI segmentation with quantifiable uncertainty.
• This approach enhances the clinical applicability of deep learning in cardiology by providing trustworthy segmentation results.
• The framework's ability to assess uncertainty aids clinicians in identifying and managing potentially erroneous segmentations.