Deep Learning Super-Resolution CT Versus Ultra-High-Resolution CT: A Direct Comparison in Cadaveric Human Lungs

Deep learning-based high-resolution computed tomography (CT) of the lung parenchyma has not been directly compared with ultra-high-resolution CT (UHR-CT). We compared image quality between deep learning-based super-resolution CT and UHR-CT using identical cadaveric human lungs.

This retrospective study included 10 cadaveric human lungs scanned using conventional CT and UHR-CT at two dose levels (9 mGy, 2 mGy). Five reconstruction methods were evaluated, including deep learning-based synthetic super-resolution CT (CCT-1024-PIQE) and UHR-CT-derived high-resolution images reconstructed using a deep learning-based denoising algorithm (UHR-1024-AiCE). Two radiologists independently assessed the bronchioles, seven abnormal findings, and overall image quality using a 5-point scale. The objective image noise was also quantified. Statistical comparisons were performed using the Wilcoxon signed-rank test with Benjamini-Hochberg correction.

At 9 mGy, no significant differences were observed in the bronchioles, whereas UHR-1024-AiCE demonstrated higher scores than CCT-1024-PIQE for ground-glass opacities and intralobular reticulation (all P ≤ 0.007). At 2 mGy, CCT-1024-PIQE showed no significant differences from UHR-1024-AiCE for most findings. Compared with the 512-matrix reconstructions, CCT-1024-PIQE achieved consistently higher visual scores at both dose levels (most comparisons, P < 0.05). Image noise increased at 2 mGy across all methods, while PIQE-based reconstructions showed lower noise (P < 0.001).

Deep learning-based super-resolution CT showed no significant differences from UHR-CT in many pulmonary findings, particularly under low-dose conditions, while improving visualization compared to conventional 512-matrix reconstructions.