Benchmark White Matter Hyperintensity Segmentation Methods Fail on Heterogeneous Clinical MRI: A New Dataset and Deep...

Area of Science:

• Medical Imaging
• Artificial Intelligence in Medicine
• Neuroimaging

Background:

• Automated white matter hyperintensity (WMH) segmentation methods struggle with generalization across heterogeneous clinical MRI data.
• Variability in scanners, field strengths, and protocols significantly impacts segmentation accuracy.

Purpose of the Study:

• To develop and evaluate robust deep learning models for WMH segmentation using a diverse clinical dataset.
• To compare the performance of a standard nnU-Net model and a fine-tuned foundation model (MedSAM) for WMH segmentation.

Main Methods:

• A diverse dataset of 195 routine brain MRI scans from 71 scanners was curated, with manual WMH annotations.
• Two models were developed: Robust-WMH-UNet (trained nnU-Net) and Robust-WMH-SAM (fine-tuned MedSAM).
• Performance was benchmarked against existing segmentation methods using metrics like the Dice Similarity Coefficient (DSC).

Main Results:

• Benchmark methods showed poor generalization, missing small lesions and generating false positives.
• Robust-WMH-UNet achieved a superior median DSC of 0.768 with improved specificity.
• Robust-WMH-SAM demonstrated competitive performance (median DSC up to 0.750) with rapid training.

Conclusions:

• The developed clinically representative dataset is crucial for advancing robust WMH segmentation algorithms.
• Both Robust-WMH-UNet and Robust-WMH-SAM show promise for reliable WMH segmentation in clinical practice.
• These findings support the translation of advanced segmentation models into routine clinical workflows.