Plug and Play Labeling Strategies for Boosting Small Brain Lesion Segmentation.

Area of Science:

• Medical Imaging
• Neuroscience
• Computer Vision

Background:

• Accurate segmentation of small brain lesions in MRI is crucial for neurological disorder diagnosis and treatment.
• Detecting small lesions is challenging due to low contrast and limited size, impacting clinical decisions.

Purpose of the Study:

• To introduce and evaluate two novel labeling strategies, Multi-Size Labeling (MSL) and Distance-Based Labeling (DBL), for enhancing small brain lesion segmentation.
• To demonstrate the generalizability and effectiveness of MSL and DBL across different datasets and segmentation architectures.

Main Methods:

• Proposed MSL strategy groups lesions by volume for size-aware learning.
• Proposed DBL strategy emphasizes lesion boundaries to improve structural sensitivity.
• Integrated MSL and DBL into existing segmentation networks and evaluated on stroke and multiple sclerosis lesion datasets (ATLAS v2.0, MSLesSeg).

Main Results:

• Achieved superior performance on the ATLAS v2.0 dataset, outperforming the previous top method in Dice, F1, precision, and recall.
• Secured the highest Dice score on the MSLesSeg dataset, ranking first among 16 international teams.
• Demonstrated consistent improvements across various segmentation models, including attention-based and mamba-based architectures.

Conclusions:

• MSL and DBL are effective and generalizable strategies for improving small brain lesion segmentation in MRI.
• These methods offer a robust solution for enhancing diagnostic accuracy and clinical decision-making in neurological disorders.
• The proposed labeling strategies provide a significant advancement in medical image analysis for neuroimaging applications.