PULASki: Learning inter-rater variability using statistical distances to improve probabilistic segmentation.

Area of Science:

• Medical Imaging
• Computational Biology
• Machine Learning

Background:

• Supervised learning for medical image segmentation struggles with annotation variability, limited data, and class imbalance.
• Existing methods can produce imprecise segmentations and lack uncertainty quantification, hindering clinical utility.

Purpose of the Study:

• To introduce PULASki, a computationally efficient generative method for biomedical image segmentation.
• To address challenges in expert annotation variability, small datasets, and class imbalance.
• To improve segmentation precision and anatomical plausibility for clinical applications.

Main Methods:

• PULASki utilizes a conditional variational autoencoder with an improved loss function based on statistical distances (Probabilistic UNet).
• The method was evaluated on intracranial vessel and multiple sclerosis (MS) lesion segmentation tasks using challenging, class-imbalanced datasets.
• A comparative study explored 3D patch-based segmentation against traditional 2D slice-based approaches.

Main Results:

• PULASki significantly outperformed four well-established baseline methods across quantitative and qualitative metrics (p < 0.05).
• The method demonstrated superior performance in learning the conditional decoder, especially for class-imbalanced problems.
• 3D patch-based segmentation yielded more anatomically plausible results than 2D slice-based methods, particularly for vessel segmentation.

Conclusions:

• PULASki offers an effective and efficient solution for biomedical image segmentation, particularly in scenarios with limited or variable annotations.
• The method's ability to capture annotation variability and improve segmentation accuracy has significant implications for clinical decision-making and treatment planning.
• PULASki is versatile for multi-label segmentation and downstream applications like hemodynamic modeling.