Deep Learning-Based Iodine Contrast Augmentation for Suboptimally Enhanced CT Pulmonary Angiography: Implications for Pulmonary Embolism Diagnosis

Abstract

Background/Objectives: This study aimed to assess the impact of a deep learning-based iodine contrast augmentation (DLCA) algorithm on image quality and diagnostic performance for pulmonary embolism (PE) detection in suboptimally enhanced CT pulmonary angiography (CTPA). Methods: We retrospectively included 103 suboptimal CTPA cases performed between May 2020 and March 2025. Image quality (attenuation, noise, SNR, and CNR) was compared between original and DLCA-processed images. Diagnostic performance for PE detection was assessed per segment, with and without DLCA processing. Results: DLCA increased pulmonary artery opacification by 57.7% and reduced noise by 56.7%, significantly improving SNR (13.2 -> 47.5) and CNR (8.7 -> 37.2; both p < 0.001). Incorporation of DLCA-processed images improved diagnostic accuracy for overall (AUC: 0.874/0.845 -> 0.958/0.938), central (0.939/0.895 -> 0.987/0.972), and peripheral (0.824/0.807 -> 0.935/0.912) PE detection (all p <= 0.003). In suboptimal CTPA, a pulmonary artery attenuation threshold of 130 HU was identified, above which DLCA processing significantly improved PE detection accuracy compared with original images in both readers (p < 0.001). Conclusions: DLCA processing in suboptimal CTPA significantly enhances image quality and diagnostic accuracy for PE detection, providing a promising strategy to optimize scans without additional contrast or radiation.