Multi-class medical image segmentation using one-vs-rest graph cuts and majority voting.

Area of Science:

• Medical image analysis
• Computer-assisted diagnosis
• Computational pathology

Background:

• Semi-automatic image segmentation is crucial for clinical applications, requiring expert oversight.
• Simultaneous multi-class segmentation methods are often complex and difficult to implement clinically.

Purpose of the Study:

• To develop an efficient one-vs-rest graph cut approach for semi-automatic multi-class image segmentation.
• To address the complexity limitations of traditional multi-class segmentation algorithms.

Main Methods:

• Constructing one-vs-rest graphs for each tissue class to infer a conditional random field (CRF).
• Minimizing CRF energy using regional and boundary class probabilities from random forests for one-vs-rest segmentation.
• Fusing individual segmentations via majority voting for the final output.
• Evaluating performance on brain tumor segmentation using the 2013 MICCAI dataset.

Main Results:

• The proposed method achieved a mean Dice score of 0.83 for whole tumor segmentation.
• Outperformed alpha-beta swap (0.80) and fully connected CRF (FCCRF) (0.79) methods.
• Demonstrated a five-fold performance improvement over the alpha-beta swap method.

Conclusions:

• The one-vs-rest graph cut approach offers linear complexity growth with the number of classes, enhancing clinical applicability.
• The method's reliance on probabilistic CRF, estimable via machine learning, allows flexibility.
• Suitable for both online semi-automatic and offline automatic segmentation in clinical settings.