Investigation of an efficient multi-modal convolutional neural network for multiple sclerosis lesion detection

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Neuroscience

Background:

• Multiple Sclerosis (MS) lesion segmentation in MRI is crucial for diagnosis and monitoring.
• Accurate segmentation of MS lesions from multi-modal MRI (mmMRI) presents significant challenges.
• Existing automated methods may lack speed, precision, or adaptability across different scanners.

Purpose of the Study:

• To develop and evaluate an automated 2D machine learning approach for fast and precise segmentation of MS lesions from mmMRI.
• To leverage a U-Net like convolutional neural network (CNN) architecture for slice-based segmentation.
• To assess the performance and robustness of the proposed method on public datasets.

Main Methods:

• An automated 2D U-Net like CNN was designed for slice-based segmentation of brain MRI volumes.
• Individual MRI modalities were processed in separate downsampling branches without weight sharing.
• Skip connections and multi-scale feature fusion/upsampling blocks were incorporated to enhance feature utilization.

Main Results:

• The proposed 2D CNN architecture achieved top-tier performance in the ISBI 2015 MS lesion segmentation challenge.
• The method outperformed state-of-the-art 3D approaches without requiring post-processing.
• The approach demonstrated robustness against scanner variability and adaptability to different scanners.

Conclusions:

• The developed automated 2D machine learning method offers fast and precise MS lesion segmentation from mmMRI.
• The U-Net based CNN is a viable and high-performing alternative to existing segmentation techniques.
• The system is efficient, robust, and deployable on standard hardware, facilitating wider clinical application.