Incorporating normal periventricular changes for enhanced pathological white matter hyperintensity segmentation: on...

Area of Science:

• Neuroimaging
• Artificial Intelligence in Medicine
• Cerebrovascular Diseases

Background:

• White matter hyperintensities (WMH) on FLAIR MRI are key biomarkers for cerebrovascular pathology, linked to cognitive decline and stroke.
• Current automated segmentation methods struggle to differentiate pathological WMH from normal age-related changes, leading to high false-positive rates and limited clinical use.

Purpose of the Study:

• To investigate if incorporating normal WMH as a distinct class in deep learning training improves pathological WMH detection compared to binary segmentation.
• To evaluate the performance of different deep learning architectures under multiclass versus binary segmentation paradigms.

Main Methods:

• Four deep learning architectures (U-Net, Attention U-Net, DeepLabV3Plus, Trans-U-Net) were trained using two paradigms: binary (background vs. pathological WMH) and multiclass (background, normal WMH, pathological WMH).
• The study utilized 2,750 FLAIR MRI images from 115 patients with neurodegenerative diseases, with expert radiological annotations.
• Performance was statistically evaluated using paired comparative analysis and Cohen's d for effect size.

Main Results:

• The U-Net architecture showed the most significant improvement with multiclass training, enhancing the Dice coefficient by 0.271 (0.768 vs. 0.497) and improving Hausdorff distance (11.5 vs. 13.4, p < 0.0001).
• All evaluated architectures demonstrated statistically significant benefits and medium practical effects (Cohen's d = 0.44-0.57) from multiclass training.
• Convolutional neural network-based architectures exhibited more stable training and greater performance gains than transformer-based models in this dataset.

Conclusions:

• Multiclass training, explicitly including normal WMH, substantially enhances the identification of pathological WMH.
• This approach offers a robust framework for improving automated neuroimaging diagnostics, maintaining clinical practicality.
• The findings support the clinical utility of advanced deep learning techniques for accurate WMH segmentation in neurodegenerative disease assessment.