سامانه همایش‌ها TSTA | ثبت‌نام و ارسال مقاله کنفرانس‌ها و کنگره‌های ملی

Home / سی و دومین کنفرانس ملی و دهمین کنفرانس بین المللی مهندسی زیست پزشکی ایران

Leveraging Normal White Matter Hyperintensity Context for Enhanced Pathological Segmentation via Multi-Class Deep Learning

Authors :

Mahdi Bashiri Bawil¹ Mousa Shamsi² Ali Fahmi Jafargholkhanloo³ Abolhassan Shakeri Bavil⁴

1- Tabriz University of Technology (Sahand) 2- Tabriz University of Technology (Sahand) 3- University of Mohaghegh Ardabili 4- Department of Radiology, Imam Reza Hospital Tabriz University of Medical Sciences Tabriz, Iran

Keywords :

White matter hyperintensities (WMH)،deep learning،medical image segmentation،FLAIR MRI،multi-class classification،U-Net،pathological segmentation،neuroimaging

Abstract :

White matter hyperintensities (WMHs) on FLAIR MRI are critical indicators of cerebrovascular dysfunction associated with elevated risks of stroke, dementia, and death. Current automated segmentation methods suffer from false positive detection in periventricular regions, failing to distinguish normal or aging-related hyperintensities from pathologically significant lesions, which reduces clinical applicability and diagnostic accuracy. This study investigates whether training deep learning models to explicitly differentiate between normal and abnormal WMH improves pathological WMH segmentation performance compared to traditional binary approaches. Four state-of-the-art architectures (U-Net, Attention U-Net, DeepLabV3Plus, Trans-U-Net) were evaluated across two training scenarios using 1,974 FLAIR images from 100 MS patients with expert-annotated ground truths. Scenario 1 employed binary training (background vs abnormal WMH), while Scenario 2 utilized three-class training (background, normal WMH, abnormal WMH). Statistical analysis included paired t-tests and Cohen's d effect size calculations. U-Net achieved the most substantial improvement in Scenario 2 with 55.6% increase in Dice coefficient (0.693 vs 0.443) and 131% precision enhancement (p < 0.0001, Cohen's d = 0.971). Traditional CNN-based architectures demonstrated larger effect sizes than transformer-based models. The three-class training approach significantly enhances pathological WMH segmentation while maintaining clinical feasibility, providing a validated framework for improving automated neuroimaging tools' diagnostic utility.