Performance evaluation of deep learning image reconstruction algorithm for dual-energy spectral CT imaging: A phantom...

Area of Science:

• Medical Imaging
• Radiology
• Computational Imaging

Background:

• Dual-energy spectral CT (DEsCT) provides material-specific information.
• Image reconstruction algorithms significantly impact DEsCT image quality.
• Deep learning image reconstruction (DLIR) is a novel approach for CT image processing.

Purpose of the Study:

• To evaluate the performance of DLIR in DEsCT.
• To compare DLIR with filtered-back-projection (FBP) and adaptive statistical iterative reconstruction-V (ASIR-V) algorithms.
• To assess the impact of radiation dose and energy levels on reconstruction performance.

Main Methods:

• An ACR464 phantom was scanned using DEsCT at varying radiation doses.
• Virtual monochromatic images were reconstructed at different energy levels using FBP, ASIR-V, and DLIR (DLIR-L, DLIR-M, DLIR-H).
• Noise power spectrum (NPS), task-based transfer function (TTF), and detectability index (d') were computed and compared.

Main Results:

• DLIR demonstrated better noise containment, especially at lower keV, compared to FBP and ASIR-V.
• DLIR exhibited higher TTF(50%) for soft tissue-like materials.
• DLIR-M and DLIR-H achieved higher detectability (d') across all dose and energy levels, outperforming ASIR-V even at lower radiation doses.

Conclusions:

• DLIR significantly improves noise containment and detectability in DEsCT images.
• DLIR-H offers the lowest noise and highest detectability across all tested conditions.
• DLIR-M and DLIR-H show potential for substantial radiation dose reduction in DEsCT while enhancing image quality.