Multiple Sclerosis Lesions Segmentation Using Attention-Based CNNs in FLAIR Images.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Neuroscience

Background:

• Multiple Sclerosis (MS) is a central nervous system autoimmune disease causing demyelination and lesions.
• Magnetic Resonance Imaging (MRI) is crucial for MS diagnosis and tracking.
• Current multimodality lesion segmentation methods are costly, time-consuming, and less user-friendly.

Purpose of the Study:

• To develop an accurate, cost-effective, and efficient method for segmenting Multiple Sclerosis (MS) lesions.
• To utilize a single MRI modality (FLAIR) for lesion segmentation, reducing complexity and resource requirements.

Main Methods:

• A patch-based Convolutional Neural Network (CNN) inspired by 3D-ResNet and a spatial-channel attention module was designed.
• The method involves Contrast-Limited Adaptive Histogram Equalization (CLAHE), edge extraction, and patch-based CNN processing.
• The architecture incorporates convolution, deconvolution, and an SCA-VoxRes attention module.

Main Results:

• The proposed single-modality FLAIR approach significantly outperformed existing methods in Dice similarity and Absolute Volume Difference on the ISIB challenge dataset.
• Experimental results demonstrate superior performance compared to previous studies using the same dataset.
• The method achieves accurate MS lesion segmentation using only one imaging modality.

Conclusions:

• An automated, efficient, and accurate method for MS lesion segmentation has been developed.
• The proposed attention-based CNN architecture effectively segments lesions using minimal input data.
• This approach offers a promising alternative to complex, multi-modal segmentation techniques for MS management.