Optimization of hypovascular liver lesion detectability in dual-energy CT using deep learning image reconstruction: a phantom study for potential iodine dose reduction

OBJECTIVE: To determine the optimal low-keV level using deep learning image reconstruction (DLIR) that maximizes lesion detectability, and to assess the potential for iodinated contrast media (ICM) reduction based on detectability improvements across varying patient body habitus. MATERIALS AND METHODS: An abdominal phantom was scanned using a standard thoraco-abdomino-pelvic dual-energy computed tomography (DECT) protocol during the portal venous phase, with three rings inserted simulating different body habitus. Virtual monoenergetic images (VMI) were reconstructed from 40 to 70 keV in 10 keV increments using adaptive statistical iterative reconstruction-V (ASIR-V) 50% and high-strength DLIR (DLIR-H). Contrast enhancement was quantified, spatial resolution was evaluated with the task-based transfer function, and noise characteristics were analyzed using the noise power spectrum. Low-contrast lesion detectability (5-10 mm) was assessed using an anthropomorphic model observer. RESULTS: Compared to ASIR-V, DLIR-H provided equivalent contrast, reduced image noise, and improved spatial resolution. All lesion sizes with DLIR-H were technically detectable under all conditions. The reconstruction at 40 keV demonstrated the highest detectability of hypovascular lesions under all conditions. However, a decrease in detectability was observed in the large phantom relative to the small and medium phantoms, resulting in a reduced theoretical potential for iodine dose reduction. The theoretical potential for iodine dose reduction using 40 keV with DLIR-H is at least 31.3% based on the phantom-based model. CONCLUSION: Under phantom conditions, 40 keV with DLIR-H shows superior detectability of hypovascular lesions under all conditions, suggesting the theoretical possibility of reducing iodine load by up to 31.3%, based on modeled detectability performance. RELEVANCE STATEMENT: Based on a phantom-derived model, the combination of 40-keV VMI reconstruction with DLIR-H suggests the potential for more than 30% ICM reduction in oncologic body CT, a finding that warrants confirmation in clinical studies. KEY POINTS: Based on a phantom-derived model 40 keV VMI with DLIR-H achieved the highest detectability of hypovascular liver lesions. This approach enabled a 31.3% ICM volume reduction. Larger body habitus limits ICM volume reduction optimization margins.