Sampling image segmentations for uncertainty quantification.

Area of Science:

• Medical Image Analysis
• Computational Imaging
• Machine Learning

Background:

• Accurate image segmentation is crucial for medical diagnosis and treatment planning.
• Generating diverse segmentation samples is essential for uncertainty quantification.
• Existing methods may lack computational efficiency or spatial coherence.

Purpose of the Study:

• To develop an automated method for producing plausible image segmentation samples from a single expert segmentation.
• To enable uncertainty quantification in medical image analysis and radiotherapy planning.
• To provide a computationally efficient and visually plausible approach.

Main Methods:

• Defining a probability distribution of image segmentation boundaries using Gaussian processes.
• Utilizing supervoxels for sampling segmentations with non-stationary covariance functions.
• Extending the method to handle multiple structures, under/over segmentation, and excluded regions.

Main Results:

• The proposed method generates spatially coherent and visually plausible segmentation samples.
• Sample variability is controlled by a user-defined parameter, potentially correlated with Dice's coefficient.
• The approach was validated by comparing generated samples with manual clinical segmentations of brain tumors.

Conclusions:

• The developed method offers an efficient way to generate diverse segmentation samples.
• This technique is valuable for uncertainty quantification in radiotherapy planning, including clinical target volumes and organs at risk.
• The approach demonstrates applicability in real-world clinical scenarios.