Segmentation of multicorrelated images with copula models and conditionally random fields

Area of Science:

• Medical Imaging
• Computer Vision
• Statistical Modeling

Background:

• Multisource medical images offer complementary information (e.g., MRI T1/T2) but can suffer from redundancy and correlation.
• Efficiently fusing multisource data without reinforcing redundancy is a key challenge in medical image analysis.
• Previous work introduced copula models within Hidden Markov Fields (HMF) for multisource image segmentation.

Purpose of the Study:

• To propose and evaluate a novel method for segmenting statistically correlated multisource images.
• To efficiently fuse complementary information from multiple sources while mitigating redundancy.
• To compare the performance of copula-based Hidden Markov Fields (HMF) and Conditional Random Fields (CRF) for multisource image segmentation.

Main Methods:

• Developed a segmentation method incorporating copula, a functional dependency measure, within Conditional Random Fields (CRF).
• CRF models the conditional distribution of hidden states given observations using an energy function with two terms: intensity similarity and spatial proximity.
• Compared the proposed CRF method with HMF and state-of-the-art supervised (CNNs) and unsupervised (hierarchical MRF) methods using simulated and real BRATS 2013 data.

Main Results:

• The copula significantly improved segmentation results in both HMF and CRF models compared to using individual image sources.
• The proposed unsupervised CRF method achieved performance comparable to decision trees on synthetic data.
• For real medical images, the unsupervised CRF method demonstrated results similar to supervised Convolutional Neural Networks (CNNs).

Conclusions:

• Copula-based statistical methods (HMF and CRF) are effective for segmenting correlated multisource medical images.
• The incorporation of copulas demonstrably enhances segmentation accuracy by effectively modeling dependencies between image sources.
• The proposed unsupervised CRF method offers a competitive alternative to supervised approaches for multisource image segmentation.