Improved Reliability of Automated ASPECTS Evaluation Using Iterative Model Reconstruction from Head CT Scans

Area of Science:

• Radiology and Medical Imaging
• Neurology
• Artificial Intelligence in Healthcare

Background:

• Iterative model reconstruction (IMR) enhances computed tomography (CT) image quality over hybrid iterative reconstruction (HIR).
• High-image quality is crucial for accurate Alberta Stroke Program Early CT Score (ASPECTS) assessment in acute stroke.

Purpose of the Study:

• To evaluate the reliability of ASPECTS assessments by human raters and automated software.
• To compare ASPECTS reliability using both hybrid iterative reconstruction (HIR) and iterative model reconstruction (IMR) CT image data.

Main Methods:

• ASPECTS was assessed by three neuroradiologists and automated software on CT scans from 47 patients with large vessel occlusions.
• Assessments were performed using both HIR (5 mm slice thickness) and IMR (5 mm and 0.9 mm slice thickness) reconstructions.
• Agreement with a consensus reading (including MRI) was measured using square-weighted kappa (κ).

Main Results:

• Human raters achieved moderate to almost perfect agreement (κ = .557–.845) with consensus.
• Attending neuroradiologist showed highest agreement with HIR (κ = .845); residents showed substantial agreement across reconstructions.
• Software agreement with consensus was substantial to almost perfect, improving with IMR and 0.9 mm slice thickness (κ = .751–.814).

Conclusions:

• Both human and software-based ASPECTS ratings demonstrate good reliability across different CT reconstruction methods.
• Human rater performance correlated with their experience (attending favored HIR).
• Automated software demonstrated superior performance with IMR, particularly at 0.9 mm slice thickness, due to improved image contrast and resolution.