Subject-Based Transfer Learning in Longitudinal Multiple Sclerosis Lesion Segmentation.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Neurology

Background:

• Accurate lesion segmentation in magnetic resonance imaging (MRI) is crucial for analyzing longitudinal data in multiple sclerosis (MS).
• Existing methods may lack consistency for tracking disease progression over time.

Purpose of the Study:

• To introduce two novel transfer learning pipelines for enhanced lesion segmentation in longitudinal MS datasets.
• To improve the accuracy and consistency of automated lesion segmentation for MS patients.

Main Methods:

• Developed subject-specific transfer learning approaches to fine-tune deep learning models using initial scans.
• Applied these models to improve segmentation performance on subsequent scans for the same individual.
• Utilized linear mixed-effects (LME) models to assess statistical power in analyzing longitudinal lesion volume changes.

Main Results:

• The proposed methods demonstrated statistically significant improvements over traditional transfer learning, with a 2% Dice score, 6% sensitivity increase, and 16% reduction in average volumetric difference.
• Visual analysis confirmed enhanced segmentation quality on both public and in-house datasets.
• Subject-wise transfer learning increased statistical power for measuring longitudinal lesion volume changes.

Conclusions:

• The novel transfer learning pipelines effectively improve lesion segmentation in longitudinal MS studies.
• These advancements reduce the need for manual correction of automated segmentations, streamlining data analysis.
• The methodology enhances the reliability of lesion volume measurements for tracking MS progression.