Deep learning image reconstruction algorithm reduces image noise while alters radiomics features in dual-energy CT in...

Area of Science:

• Medical Imaging
• Radiology
• Artificial Intelligence in Medicine

Background:

• Dual-energy CT (DECT) is crucial for material decomposition and tissue characterization.
• Image reconstruction algorithms significantly impact DECT image quality and downstream analyses like radiomics.
• Deep learning image reconstruction (DLIR) offers potential improvements over conventional iterative reconstruction (IR).

Purpose of the Study:

• To compare image quality between DLIR and conventional IR algorithms in DECT.
• To evaluate the impact of DLIR versus IR on radiomics feature robustness and reproducibility.
• To identify reliable radiomics features from DECT images reconstructed with different algorithms.

Main Methods:

• A phantom was scanned on seven DECT scanners using standard abdominal-pelvis protocols.
• Raw DECT data were reconstructed using conventional IR (ASIR-V) and DLIR.
• Image quality metrics, radiomics features (Pyradiomics), and reproducibility (Bland-Altman, ICC, CCC, CV, QCD) were assessed.

Main Results:

• DLIR significantly improved DECT image quality, showing better noise reduction and signal-to-noise ratio compared to ASIR-V.
• Most radiomics features were repeatable across scans (93.87%).
• Reproducibility was lower between DLIR and conventional IR, and across different scanners.

Conclusions:

• DLIR enhances DECT image quality but introduces differences in radiomics features compared to conventional IR.
• Despite reproducibility challenges, nine radiomics features demonstrated robustness across scanners and reconstruction methods.
• These robust features warrant further validation for clinical applications.