Incorporating normal periventricular changes for enhanced pathological white matter hyperintensity segmentation: on multiclass deep learning approaches.
Authors
Affiliations (4)
Affiliations (4)
- Biomedical Engineering Faculty, Sahand University of Technology, Tabriz, Iran. [email protected].
- Biomedical Engineering Faculty, Sahand University of Technology, Tabriz, Iran. [email protected].
- Department of Engineering Sciences, Faculty of Advanced Technologies, University of Mohaghegh Ardabili, Namin, Iran.
- Radiology Department, Tabriz University of Medical Sciences, Tabriz, Iran.
Abstract
White matter hyperintensities (WMH) detected on FLAIR MRI sequences serve as important biomarkers for cerebrovascular pathology, correlating with increased risks of cognitive decline, stroke, and demyelination. Contemporary automated segmentation approaches face substantial challenges in distinguishing pathological lesions from normal age-related periventricular hyperintensities, resulting in elevated false-positive rates that limit clinical utility. This investigation examines whether incorporating normal WMH as an explicit class during deep learning model training enhances pathological lesion detection compared to conventional binary segmentation approaches. We evaluated four established architectures (U-Net, Attention U-Net, DeepLabV3Plus, and Trans-U-Net) across two training paradigms using 2,750 FLAIR images from 115 patients with neurodegenerative diseases, sourced from local and public datasets with expert radiological annotations. The first paradigm employed traditional binary classification (background versus pathological WMH), while the second utilized multiclass classification incorporating normal periventricular hyperintensities as a distinct category. Statistical evaluation included paired comparative analysis and effect size quantification using Cohen's d. The U-Net architecture demonstrated the most pronounced improvement with the multiclass approach, achieving 0.271 improvement in Dice coefficient (0.768 versus 0.497) and 1.9 improvement in Hausdorff distance (11.5 vs 13.4) (p < 0.0001, Cohen's d = 0.5643). All architectures demonstrated medium practical effects (d = 0.44-0.57) beyond statistical significance. In the present data regime, convolutional neural network-based architectures demonstrated more stable training dynamics and larger performance improvements compared to the transformer-based models, though all architectures showed statistically significant benefits from multiclass training. The multiclass training methodology substantially improves pathological WMH identification while preserving clinical practicality, offering a robust framework for enhancing automated neuroimaging diagnostic capabilities.Trial Registration Number Tabriz University of Medical Sciences Research Ethics Committee (IR.TBZMED.REC.1402.902).