Deep-Learning Generated Synthetic Material Decomposition Images Based on Single-Energy CT to Differentiate Intracranial Hemorrhage and Contrast Staining Within 24 Hours After Endovascular Thrombectomy

Abstract

Aims: To develop a transformer-based generative adversarial network (trans-GAN) that can generate synthetic material decomposition images from single-energy CT (SECT) for real-time detection of intracranial hemorrhage (ICH) after endovascular thrombectomy.

Materials: We retrospectively collected data from two hospitals, consisting of 237 dual-energy CT (DECT) scans, including matched iodine overlay maps, virtual noncontrast, and simulated SECT images. These scans were randomly divided into a training set (n = 190) and an internal validation set (n = 47) in a 4:1 ratio based on the proportion of ICH. Additionally, 26 SECT scans were included as an external validation set. We compared our trans-GAN with state-of-the-art generation methods using several physical metrics of the generated images and evaluated the diagnostic efficacy of the generated images for differentiating ICH from contrast staining.

Results: In comparison with other generation methods, the images generated by trans-GAN exhibited superior quantitative performance. Meanwhile, in terms of ICH detection, the use of generated images from both the internal and external validation sets resulted in a higher area under the receiver operating characteristic curve (0.88 vs. 0.68 and 0.69 vs. 0.54, respectively) and kappa values (0.83 vs. 0.56 and 0.51 vs. 0.31, respectively) compared with input SECT images.

Conclusion: Our proposed trans-GAN provides a new approach based on SECT for real-time differentiation of ICH and contrast staining in hospitals without DECT conditions.

Keywords: deep learning; dual‐energy CT; endovascular thrombectomy; generative adversarial networks; hemorrhagic transformation; material decomposition images; postinterventional cerebral hyperdensity; stroke.