A deep learning algorithm for white matter hyperintensity lesion detection and segmentation.

Area of Science:

• Neuroimaging
• Artificial Intelligence in Medicine
• Computational Pathology

Background:

• White matter hyperintensity (WMHI) lesions are key indicators of neurological conditions involving inflammation and vascular issues.
• Current automated quantification tools often lack generalizability across different diseases and MRI acquisition protocols.
• Accurate WMHI quantification is crucial for patient management and disease monitoring.

Purpose of the Study:

• To develop and validate a robust algorithm for automated white matter lesion (WML) quantification.
• To ensure the algorithm's applicability to heterogeneous MRI data, including various disease types and scanning parameters.
• To create a tool that can be widely used in clinical settings.

Main Methods:

• A deep learning approach named DeepWML was developed for automated segmentation of WMLs.
• The method was trained and evaluated on multicentre FLAIR MRI data from 507 patients with three distinct white matter diseases.
• Performance was assessed using Dice similarity, sensitivity, and precision against manual delineations.

Main Results:

• DeepWML achieved a median Dice similarity coefficient of 0.78 across diverse datasets and disease types.
• Median sensitivity and precision were 0.84 and 0.81, respectively, demonstrating high accuracy.
• The algorithm's performance was positively correlated with larger white matter lesion volumes.

Conclusions:

• DeepWML demonstrates successful application and validation across a wide range of MRI data and white matter diseases.
• The developed algorithm has the potential to significantly enhance the efficiency and accuracy of clinical WML delineation workflows.
• This automated tool offers a promising solution for standardized WMHI quantification in clinical practice.