Deep Unfolding Segmentation Network for Under-Sampled Magnetic Resonance Images.

Area of Science:

• Medical imaging
• Artificial intelligence
• Image processing

Background:

• Magnetic Resonance (MR) image segmentation is crucial for disease diagnosis.
• Clinical MR images are often reconstructed from under-sampled k-space data, leading to artifacts and reduced segmentation accuracy.
• Existing segmentation methods fail to address challenges posed by under-sampled MR data.

Purpose of the Study:

• To propose an end-to-end deep unfolding framework for segmenting lesions or organs directly from under-sampled MR k-space data.
• To develop a novel model combining compressive sensing-based reconstruction and level-set segmentation.
• To enhance segmentation performance by incorporating an L0 norm for image smoothing and a boundary loss function.

Main Methods:

• Developed a deep unfolding segmentation network (DUSNet) by unfolding an iterative algorithm derived from the Augmented Lagrangian Method.
• Integrated compressive sensing reconstruction with level-set segmentation, utilizing an L0 norm to preserve edges and boundaries.
• Introduced a boundary loss function to improve edge detail capture and impose geometric constraints.

Main Results:

• The proposed DUSNet effectively segments target regions from under-sampled k-space data through end-to-end training.
• The L0 norm regularization and boundary loss function significantly boosted downstream segmentation performance.
• Comprehensive experiments confirmed that DUSNet achieves superior segmentation accuracy compared to state-of-the-art methods for under-sampled MR images.

Conclusions:

• DUSNet offers an effective solution for segmenting under-sampled MR images, addressing limitations of current methods.
• The framework demonstrates the potential of deep unfolding for integrating image reconstruction and segmentation tasks.
• The proposed approach achieves state-of-the-art performance, paving the way for improved clinical MR image analysis.