Task based evaluation of sparse view CT reconstruction techniques for intracranial hemorrhage diagnosis using an AI...

Area of Science:

• Medical Imaging
• Artificial Intelligence in Radiology
• Computational Imaging

Background:

• Sparse-view computed tomography (CT) reduces radiation dose but introduces artifacts with traditional reconstruction algorithms like Filtered Backprojection (FBP) and Model-Based Iterative Reconstruction (MBIR).
• Task-based evaluations of Deep Learning Reconstruction (DLR) for sparse-view CT are limited, hindering its clinical adoption.

Purpose of the Study:

• To evaluate the diagnostic accuracy of FBP, MBIR, and DLR for intracranial hemorrhage detection and classification using an Artificial Intelligence (AI) observer in sparse-view CT.
• To assess the image quality and diagnostic utility of different reconstruction methods in a cost-effective manner.

Main Methods:

• Trained an AI observer model on a public brain CT dataset with labeled intracranial hemorrhages.
• Simulated sparse-view CT data and reconstructed images using FBP, MBIR, and DLR.
• Assessed image quality using PSNR, SSIM, and LPIPS; evaluated diagnostic utility via ROC analysis and AUC.

Main Results:

• DLR demonstrated superior image quality metrics (PSNR, SSIM, LPIPS) over FBP and MBIR, with reduced noise and artifacts.
• The AI observer achieved the highest classification accuracy with DLR reconstructions.
• FBP showed higher task-based accuracy than MBIR, underscoring the importance of task-specific evaluations.

Conclusions:

• DLR offers an effective balance of artifact reduction and anatomical detail for sparse-view CT brain imaging.
• AI observer models provide a viable and cost-effective alternative for evaluating CT reconstruction techniques.
• This study validates DLR's potential for improving diagnostic performance in low-dose sparse-view CT applications.