Automatic Virtual Contrast-Enhanced CT Synthesis Using Dual-Energy CT and Residual U-Net with Attention Module for Detecting Pulmonary Hilar Lymphadenopathy.

Rationale And Objectives: To propose an automatic virtual contrast-enhanced chest computed tomography (CT) synthesis using dual-energy CT and a Residual U-Net with an attention module to detect clinically significant hilar lymphadenopathy.

Materials And Methods: We conducted a retrospective analysis of 2082 patients who underwent dual-energy chest CT scans. Our approach utilized a Residual U-Net combined with a Convolutional Block Attention Module (CBAM) to transform non-contrast CT images into virtual contrast-enhanced CT images. We evaluated the effectiveness of our method through quantitative and qualitative analyses and an observer study involving thoracic radiologists, focusing on the detection of significant hilar lymph nodes.

Results: Our method achieved an average peak signal-to-noise ratio of 25.082, a structural similarity index of 0.833, and mutual information of 1.568. The mean absolute error, mean squared error, and root mean squared error were reported as 0.040, 0.023, and 0.102, respectively. Compared to other methods, our proposed approach demonstrated superior performance across all evaluation metrics. In the observer study, our method exhibited a higher diagnostic accuracy for detecting hilar lymphadenopathy (69.2%) compared to the Residual U-Net-based GAN with CBAM (53.7%).

Conclusion: The integration of dual-energy computed tomography with a Residual U-Net framework augmented by CBAM presents a highly effective technique for generating synthetic contrast-enhanced chest CT images. This novel approach significantly enhances the detection of clinically significant hilar lymphadenopathy, underscoring its potential clinical utility.