Modeling Aleatoric Uncertainty in Cardiac MRI Segmentation: Probabilistic Detection and Contour Regression.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Biomedical Engineering

Background:

• Accurate cardiac MRI segmentation is crucial for assessing cardiac function via biomarkers like ejection fraction.
• Current AI methods excel in average segmentation accuracy but often lack precise, calibrated biomarkers for individual patient diagnosis.
• Biomarker uncertainty and calibration are underexplored, despite being critical for reliable clinical decision-making.

Purpose of the Study:

• To develop a probabilistic segmentation framework for cardiac MRI that explicitly models aleatoric uncertainty.
• To improve the calibration of biomarkers, such as ejection fraction, by accurately reflecting true variability.
• To enhance the reliability of cardiac function assessment in individual patients.

Main Methods:

• Proposed a probabilistic segmentation framework to model aleatoric uncertainty in cardiac MRI.
• Disentangled detection uncertainty (basal/apical slice ambiguity) using objectness probabilities.
• Modeled contour uncertainty (boundary delineation variability) via mean-variance regression of elliptic Fourier descriptors.

Main Results:

• The framework produced more informative and better-calibrated confidence estimates for ejection fraction.
• Propagated segmentation uncertainties to derived biomarkers, enhancing their reliability.
• Demonstrated improved biomarker reliability compared to conventional pixel-wise approaches, especially in settings with annotation ambiguity.

Conclusions:

• The proposed probabilistic framework effectively models uncertainty in cardiac MRI segmentation.
• Improved biomarker calibration leads to more reliable estimates of cardiac function, crucial for diagnosis.
• This approach offers enhanced diagnostic precision in clinical scenarios characterized by inherent anatomical and annotation variability.