Characterization of parotid lesions with dual-energy CT: Can virtual unenhanced images permit selective omission of true unenhanced acquisition?

Background

To investigate whether virtual unenhanced (VUE) images can permit selective omission of true unenhanced (TUE) images in multiphasic parotid dual-energy CT (DECT) without compromising attenuation accuracy or diagnostic performance.

Materials And Methods

This retrospective study included 446 lesions from 436 patients with surgically confirmed parotid lesions who underwent three-phase DECT (TUE, early and late contrast); VUE was reconstructed from both contrast phases. Agreement between VUE and TUE attenuation was assessed using Bland-Altman analysis and two one-sided tests (TOST; equivalence margin, ±10 HU), both overall and across six histopathological subgroups. Lesion detection and border clarity were compared between VUE and TUE images. XGBoost-based six-category classifiers using TUE, VUE_C1, or VUE_C2 images as input were compared to evaluate diagnostic performance. The radiation dose attributable to the TUE acquisition was estimated from the dose-length product.

Results

VUE reproduced lesion attenuation with a mean bias of less than 1 HU. TOST confirmed equivalence overall and across all six subgroups (p < 0.001), with 84.3% and 81.6% of differences within ± 10 HU. Six-category XGBoost classifiers demonstrated comparable diagnostic performance across TUE, VUE_C1, and VUE_C2 baselines. SHAP feature importance showed strong correlation across models. Background gland attenuation was overestimated on VUE by 14.8 and 16.7 HU, while lesion detection and border clarity were modestly reduced. Omitting the TUE acquisition would provide an estimated dose reduction of 2.41 ± 0.37 mSv.

Conclusion

VUE is quantitatively equivalent to TUE for parotid lesion attenuation and largely preserves internally validated diagnostic performance. These findings support reduced-dose DECT protocols with selective omission of TUE in appropriately selected patients, pending broader multicentre and multi-vendor validation.