Unsupervised quality control of segmentations based on a smoothness and intensity probabilistic model

Area of Science:

• Medical image analysis
• Computer vision
• Machine learning

Background:

• Automated quality assessment of image segmentation is crucial for clinical applications.
• Manual quality control is time-consuming for large image databases.
• Existing automated methods are often supervised, requiring trusted training data.

Purpose of the Study:

• To develop a novel unsupervised approach for assessing the quality of image segmentations.
• To enable automated detection and localization of segmentation errors.
• To identify challenging cases within large image datasets.

Main Methods:

• A generic probabilistic model is utilized for quality assessment.
• Segmentations are compared against a probabilistic model based on intensity and smoothness assumptions.
• A novel smoothness prior using label map derivative penalization is introduced, with hyperparameter estimation via variational Bayesian techniques.

Main Results:

• The method effectively ranks segmentations based on quality.
• It successfully isolates atypical segmentations automatically.
• The approach can localize potential segmentation errors within images.
• Performance prediction of segmentation algorithms is demonstrated in some cases.

Conclusions:

• The proposed unsupervised method offers an efficient solution for image segmentation quality control.
• It accurately identifies segmentation flaws and difficult cases without manual supervision.
• This technique has broad applicability in medical imaging and other computer vision tasks.