Estimating the volume of penumbra in rodents using DTI and stack-based ensemble machine learning framework.

Area of Science:

• Neuroimaging
• Biomedical Engineering
• Machine Learning in Medicine

Background:

• Investigating diffusion tensor imaging (DTI) for identifying penumbral volume (PV).
• Comparing DTI-based methods with standard gadolinium-required perfusion-diffusion mismatch (PDM).
• Utilizing a stack-based ensemble machine learning (ML) approach with enhanced explainability.

Purpose of the Study:

• To evaluate the efficacy of DTI metrics in estimating PV.
• To compare ML-estimated PV with PDM-defined PV.
• To assess the explainability and clinical relevance of the DTI-based ML model.

Main Methods:

• Middle cerebral artery occlusion in sixteen male rats.
• Penumbra identification using PDM at 30 and 90 minutes post-occlusion.
• Training five voxel-wise ML models using 11 DTI metrics and 14 distance-based features, integrated via stack-based ensemble techniques.

Main Results:

• ML-estimated PV (106.4 mL) showed high agreement with PDM-defined PV (102.0 mL).
• Volume similarity was 0.88, Pearson correlation was 0.93 (p < 0.001), and Bland-Altman bias was 2.5 mL.
• Mean diffusivity emerged as the most important feature for PV prediction.

Conclusions:

• DTI metrics and stack-based ensemble ML can accurately estimate PV, comparable to PDM.
• The DTI-based ML model estimates PV without contrast agents, benefiting patients with kidney dysfunction.
• Enhanced model explainability increases clinical relevance, warranting human studies for validation.