Longitudinal detection of new MS lesions using deep learning.

Area of Science:

• Medical Imaging Analysis
• Artificial Intelligence in Healthcare
• Neuroscience

Background:

• Detecting new multiple sclerosis (MS) lesions is vital for monitoring disease progression.
• Automating new lesion detection using machine learning is challenging due to limited annotated longitudinal data.
• Existing methods struggle with the scarcity of diverse, labeled datasets for training robust models.

Purpose of the Study:

• To develop a deep learning pipeline for automated detection and segmentation of new MS lesions.
• To overcome the limitations of insufficient annotated longitudinal data for MS lesion detection.
• To improve the accuracy and generalizability of models for identifying evolving MS lesions.

Main Methods:

• Utilized transfer learning from a model trained on single time-point segmentation tasks.
• Developed a data synthesis strategy to generate realistic longitudinal data with new lesions from single time-point scans.
• Employed data augmentation techniques to simulate MRI data diversity and increase dataset size.

Main Results:

• Each component of the proposed pipeline demonstrated an improvement in segmentation accuracy.
• The pipeline achieved state-of-the-art performance in the MSSEG2 MICCAI challenge for new MS lesion segmentation and detection.
• The combined approach of transfer learning, data synthesis, and augmentation significantly enhanced model robustness.

Conclusions:

• The proposed deep learning pipeline effectively addresses the challenge of detecting new MS lesions.
• Transfer learning and data synthesis are crucial for training accurate models with limited longitudinal data.
• This approach offers a promising solution for automated monitoring of multiple sclerosis evolution.