New multiple sclerosis lesion segmentation and detection using pre-activation U-Net.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Neuroscience

Background:

• Automated segmentation of new multiple sclerosis (MS) lesions in 3D MRI is crucial for monitoring disease progression.
• Manual lesion delineation is labor-intensive, time-consuming, and requires expertise in handling multi-modal 3D imaging data.

Purpose of the Study:

• To develop and evaluate Pre-U-Net, a 3D encoder-decoder architecture with pre-activation residual blocks, for automated segmentation and detection of new MS lesions.
• To address challenges of limited training data and class imbalance in MS lesion segmentation.

Main Methods:

• Proposed Pre-U-Net, a 3D convolutional neural network incorporating pre-activation residual blocks.
• Employed intensive data augmentation and deep supervision techniques for effective model training.
• Evaluated performance on the MSSEG-2 dataset, comparing against U-Net and Res-U-Net.

Main Results:

• Pre-U-Net achieved a Dice score of 40.3% for new lesion segmentation.
• Pre-U-Net obtained an F1 score of 48.1% for new lesion detection.
• The proposed model demonstrated superior performance compared to U-Net and Res-U-Net.

Conclusions:

• Pre-U-Net offers an effective automated solution for segmenting and detecting new MS lesions in 3D MRI data.
• The findings suggest Pre-U-Net's potential to improve the efficiency and accuracy of MS progression monitoring.
• Publicly available code and models facilitate further research and clinical application.