Comparing two deep learning algorithms for acute infarct segmentation on diffusion-weighted imaging in routine...

Area of Science:

• Artificial Intelligence in Medical Imaging
• Neurology and Stroke Imaging Analysis
• Deep Learning for Medical Image Segmentation

Background:

• Accurate infarct volume measurement on diffusion-weighted imaging (DWI) is crucial for predicting stroke outcomes and guiding endovascular thrombectomy.
• Traditional 3D U-Net deep learning models achieve high sensitivity but often produce false positives due to infarct mimics.
• There is a need for improved deep learning models that can accurately segment infarcts while minimizing false positives across various pathologies.

Purpose of the Study:

• To develop and evaluate a novel SegMamba-based deep learning model for enhanced global volumetric feature extraction in DWI infarct segmentation.
• To compare the performance of the SegMamba model against a 3D U-Net-based model using a diverse dataset of DWI hyperintense pathologies.
• To assess the diagnostic accuracy and clinical utility of both models in real-world clinical scenarios.

Main Methods:

• Two deep learning models, SegMamba and 3D U-Net, were trained on a large multicenter dataset of 10,820 DWI scans.
• Model performance was evaluated on an external test set of 2731 DWI scans and a clinical cohort of 1194 patients.
• Segmentation accuracy was quantified using Dice Similarity Coefficient (DSC) and Average Hausdorff Distance (AHD), alongside sensitivity and specificity.

Main Results:

• SegMamba and U-Net demonstrated comparable DSC in the external test set (0.786 vs 0.785).
• SegMamba significantly outperformed U-Net in AHD (1.25 mm vs 1.76 mm), indicating more precise boundary delineation.
• In the clinical dataset, SegMamba achieved higher specificity (58.80% vs 29.54%) and overall accuracy (64.07% vs 39.11%) despite slightly lower sensitivity.

Conclusions:

• Modifying the deep learning architecture to SegMamba improved classification accuracy in broader disease populations while maintaining performance in ischemic stroke cohorts.
• The SegMamba model demonstrates superior performance in differentiating true infarcts from mimics, leading to higher diagnostic accuracy in diverse clinical settings.
• Validation across varied clinical environments is essential to ensure the practical utility of deep learning models for stroke imaging analysis.