Detection of Vascular Pathologies in Non-contrast Magnetic Resonance Images for Early Screening
TL;DR
Practical recommendations for automated segmentation and future validation requirements are outlined: favor self-configuring frameworks like nnU-Net when compute resources permit, consider lightweight 2D models such as YOLOv8 for fast screening, and ensure rigorous cross-site validation before potential clinical use.
Abstract
Early, non-invasive detection of cerebrovascular pathologies is essential for improving patient triage and outcomes. Contrast-free FLAIR MRI sequences are widely available in clinical practice but pose challenges for automated analysis due to variable lesion appearance, low contrast, and inter-scanner variability. The objective of this study is to perform a comparison of three representative deep learning segmentation paradigms for ischemic lesion detection on non-contrast FLAIR MRI under identical preprocessing and evaluation conditions. This study evaluates and compares three deep learning strategies for ischemic lesion segmentation in non-contrast FLAIR images: a classical 3D U-Net, the self-configuring nnU-Net framework, and an adapted 2D YOLOv8 segmentation variant. All methods were trained and tested using the ISLES 2022 dataset (N = 250 cases) under matched preprocessing and evaluation protocols to ensure a fair comparison. Quantitative performance is reported using the Dice similarity coefficient and Intersection over Union (IoU), and trade-offs among accuracy, robustness, and inference efficiency are analyzed. The nnU-Net achieved the highest average overlap (Dice = 0.483 ± 0.284), demonstrating superior robustness across heterogeneous lesion presentations. The YOLOv8-based model provided competitive volumetric performance (Dice ≈ 0.476) while substantially reducing inference time, suggesting suitability for time-constrained clinical workflows. The classical 3D U-Net served as a manual baseline and exhibited lower mean overlap (Dice ≈ 0.294), indicating sensitivity to parameter tuning and dataset variability. Beyond numerical metrics, we qualitatively examine typical failure modes (small peripheral lesions, diffuse low-contrast areas, and spatial mismatches between masks and images) and discuss preprocessing steps that mitigate them (spatial resampling, intensity normalization, and patch-based training). Finally, the study outlines practical recommendations for automated segmentation and future validation requirements: favor self-configuring frameworks like nnU-Net when compute resources permit, consider lightweight 2D models such as YOLOv8 for fast screening, and ensure rigorous cross-site validation before potential clinical use.