Deep learning image reconstruction improves visualization of arterial phase hyperenhancement and washout appearance on dual-energy CT for hepatocellular carcinoma: a non-inferiority study.

PURPOSE: This study aimed to evaluate the clinical value of deep learning image reconstruction (DLIR)-based dual-energy CT (DECT) in improving image quality for hepatocellular carcinoma (HCC). METHODS: This single-center retrospective analysis of a prospective cohort included patients enrolled between June 2024 and July 2025. Virtual monoenergetic images (VMI) at 40, 50, 60, and 74-keV (120 kVp-like) were reconstructed using ASiR-V 50%, DLIR-H (high), and DLIR-M (medium). All combinations of energy levels and reconstruction algorithms were compared using both quantitative metrics standard deviation (SD) of liver and lesion attenuation, signal-to-noise ratio (SNR), and lesion-to-liver contrast ratio (LLR) and semi-quantitative 5-point scores (overall noise, lesion edge sharpness, and conspicuity). The optimal reconstruction combination-derived DECT image was identified and compared with MRI for major HCC features of LI-RADS 2018, including arterial phase hyperenhancement (APHE) and nonperipheral washout appearance. RESULTS: Each patient yielded 36 image sets across three phases from various combinations of energy levels and algorithms. Quantitative analysis revealed DLIR-H/50-60-keV performed best across all objective metrics (all p < 0.05); with quantitative assessment, DLIR-H/50-keV was determined as the optimal protocol, which showed non-inferiority to MRI for detecting the two major HCC features of LI-RADS 2018. CONCLUSION: DLIR significantly enhances low-energy VMI quality and the visualization of major LI-RADS 2018 features in HCC. The DLIR-H/50-keV protocol demonstrates imaging performance approaching MRI standards, representing a promising reconstructive strategy for HCC assessment, particularly in clinical scenarios where MRI access is limited.