SGD-Net

SGD-Net performs segmentation, classification, and atlas-based mapping of acute ischemic stroke (AIS) lesions from diffusion-weighted imaging (DWI) to quantify lesion characteristics and spatial distribution.

Key Features:

• U-Shape Model Architecture: Employs a U-shaped model architecture for semantic segmentation of AIS lesions in DWI, achieving Dice coefficients of 0.806–0.828 and Intersection over Union (IoU) values of 0.675–0.707.
• Binary Classification: Performs binary classification to distinguish lacunar versus non-lacunar lesions and to identify anterior versus posterior circulatory territory.
• SGD-Net Plus Enhancement: SGD-Net Plus incorporates automated segmentation and registration to align DWI lesions with T1-weighted images and brain atlases for region-wise mapping.
• Performance Metrics: A two-stage deep learning framework outperforms one-stage models, with lesion size classification accuracies of 0.867–0.956, AUROC 0.962–0.992, AUPRC 0.964–0.994, and lesion location classification accuracies of 0.860–0.930, AUROC 0.936–0.988, AUPRC 0.883–0.978.
• Quantitative Mapping: Reports lesion volume, region percentage, and lesion occupancy within selected brain regions including white matter tracts, Brodmann areas, and cytoarchitectonic areas.

Scientific Applications:

• Lesion segmentation and quantification: Produces pixel-wise segmentations and quantitative metrics for assessment of AIS lesion burden from DWI.
• Lesion classification and localization: Classifies lesion type (lacunar/non-lacunar) and determines anterior versus posterior circulation location to inform diagnostic stratification.
• Atlas-based regional analysis and research: Maps lesions to T1-weighted images and brain atlases to quantify regional lesion distribution across white matter tracts, Brodmann areas, and cytoarchitectonic areas for clinical research.

Methodology:

Uses a two-stage deep learning framework with a U-shaped semantic segmentation model followed by classification, and SGD-Net Plus adds automated registration of DWI lesion masks to T1-weighted images and brain atlases.