Attention Augmented Deep Learning-Based Dual-Energy CT Imaging Via Single-Energy CT Data
W Zhang1*, T Lv2, y chen2, B Sun1, W Zhao1, (1) Beihang University, Beijing, CN, (2) Southeast University, Nanjing, CN,
Presentations
SU-H330-IePD-F5-4 (Sunday, 7/10/2022) 3:30 PM - 4:00 PM [Eastern Time (GMT-4)]
Exhibit Hall | Forum 5
Purpose: Dual-energy CT (DECT) allows differentiation and quantification of different tissues and therefore has been used for a wide spectrum of emerging clinical applications. However, due to its cost, DECT has not been widely used yet compared to the conventional single-energy CT(SECT). Recent studies have shown it is possible to perform DECT imaging from easily available SECT via deep learning-based approaches. Here we propose an attention augmented learning-based approach to generate high quality DECT images using standard SECT data.
Methods: The attention augmented learning-based strategy employs a weight map generator and a standard U-Net with four attention augmented gates. The weight map generator extracts the features from SECT images and provides pixel-level weights for the attention augmented gates, which assist the U-Net to focus on regions of significant dual-energy differences and reduce the interference of CT noise. The performance of our approach was studied using contrast-enhanced abdomen DECT images from 22 patients. The predicted high-energy CT (HECT) images were assessed quantitatively using HU accuracy in ROIs on five different types of tissues. Comparison studies with the state-of-the-art deep learning-based DECT imaging methods are also performed.
Results: The absolute HU differences between the original HECT images generated from the DECT scan and the predicted ones are 2.6 HU, 2.4 HU, 2.0 HU, 0.4HU and 2.9 HU for the ROIs on heart, aorta, spine, liver and kidney, respectively. In addition, compared with the state-of-the-art deep learning-based DECT imaging method, we find that the distribution of ROI's CT values in HECT images predicted by our method is closer to the ground truths, indicating the merit of the proposed model.
Conclusion: The premise of the approach is that DECT can be attained without any additional measurement other than a SECT acquisition. The technique provides a simple and cost-effective solution for modern DECT clinical applications.
Funding Support, Disclosures, and Conflict of Interest: This work was supported by the National Natural Science Foundation of China (No. 12175012)
Keywords
Taxonomy
Not Applicable / None Entered.
Contact Email