Feasibility of Automated Accurate Lung Segmentation Using Deep Learning on Virtual Unenhanced Images from Gemstone Spectral CT Imaging For Pulmonary Ventilation Assessment

Objective :To evaluate the feasibility of applying deep learning-based automatic lung segmentation to virtual unenhanced (VUE) images generated from gemstone spectral imaging (GSI) CT scans. Materials and Methods :This retrospective study included patients who underwent chest CT scans. The protocol consisted of conventional true unenhanced (TUE) scans, arterial-phase (AP) GSI-enhanced scans, and venous-phase (VP) GSI-enhanced scans. VUE images were reconstructed from both enhanced phases. A deep neural network was utilized to automatically segment lung lobes and calculate total lung volume, weight and relative fractions, as well as the volume and weight fractions of regions with different ventilation functionality. Differences among the three image sets, as well as their correlations, bias, and mean absolute percentage error (MAPE), were assessed. Results :A total of 223 patients were enrolled. Statistically significant differences were observed between TUE and AP-VUE in the weight fractions of normally ventilated regions, and between TUE and both AP-VUE and VP-VUE in the weight fractions of poorly ventilated regions (all P  < 0.05). No significant differences were found among the three image types in other segmentation parameters (all P  > 0.05). The correlation coefficients of non-ventilated region weight fractions between AP-VUE and TUE, and between VP-VUE and TUE, were 0.75 and 0.76, respectively; all other correlation coefficients exceeded 0.80. Bias values for lobar and functional region volume and weight fractions ranged from − 1.62 to 1.23, while MAPE ranged from 0.00–2.37%. Compared to three-phase scanning, dual-phase enhanced scanning without TUE resulted in a 35.70% reduction in radiation dose. Conclusion :Lung parameters calculated from VUE images using deep learning-based segmentation demonstrated strong agreement with those derived from TUE, with minimal differences, high correlation, and low bias and MAPE.