Deep Attention and Graphical Neural Network for Multiple Sclerosis Lesion Segmentation From MR Imaging Sequences.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Neuroscience

Background:

• Accurate segmentation of multiple sclerosis (MS) lesions in MRI is crucial for diagnosis and treatment monitoring.
• Current deep learning methods struggle with the diverse shapes, scattered distribution, and variable numbers of MS lesions.
• Challenges include capturing distant lesions and delineating the global contours of varied lesion shapes.

Purpose of the Study:

• To develop a novel attention and graph-driven network (DAG-Net) for improved automated MS lesion segmentation.
• To address limitations in capturing scattered lesions and delineating global lesion contours.
• To enhance the interpretability and reliability of MS lesion segmentation.

Main Methods:

• Proposed DAG-Net integrates spatial correlations and global context within a unified architecture.
• A local attention coherence mechanism constructs dynamic graphs to capture spatial relationships between pixels.
• A spatial-channel attention module enhances feature representation for global contour delineation.

Main Results:

• DAG-Net demonstrated superior performance in segmenting variant and scattered MS lesions across multiple regions.
• Experiments on public (ISBI2015) and in-house datasets validated the effectiveness against state-of-the-art methods.
• The dynamic graph construction provides interpretability, enhancing the reliability of segmentation results.

Conclusions:

• DAG-Net offers a significant advancement in automated MS lesion segmentation from MR imaging.
• The proposed network effectively handles the complexities of MS lesion characteristics, including scattered distribution and variant shapes.
• The interpretability of DAG-Net contributes to trustworthy clinical applications in neuroimaging analysis.