Virtual Non Contrast Tomography Synthesis for Hepatocellular Carcinoma Patients Using Multimodality-Guided Synergistic Neural Network
J Chen1, W Li2, H Xiao3*, S Lam4, J Chen5, C Liu6, A Cheung7, J Cai8, (1)The Hong Kong Polytechnic University,Hung Hom, ,HK, (2) The Hong Kong Polytechnic University, ,,(3) The Hong Kong Polytechnic University, Hong Kong, 91, CN, (4) Duke Kunshan University, Kunshan, ,CN, (5) ,,,(6) The Hong Kong Polytechnic University, Hong Kong, Hong Kong, (7) ,,,(8) Hong Kong Polytechnic University, Hong Kong, ,CN
Presentations
TU-J430-BReP-F1-5 (Tuesday, 7/12/2022) 4:30 PM - 5:30 PM [Eastern Time (GMT-4)]
Exhibit Hall | Forum 1
Purpose: To synthesizes virtual non-contrast (VNC) computed tomography (CT) from enhanced CT images through multimodality-guided synergistic neural network (MMgSN-Net) and evaluate the validity between Predicted CT Plain and real CT Plain for hepatocellular carcinoma (HCC) patents.
Methods: 96 HCC patients who had non-contrast CT and contrast-enhanced CT with iodine were included in this study. The dateset was split to 82 and 14 for model training and testing respectively. Deformable registration was applied to register the Portal Venous Phase (CTP) and CT Plain images before model training. A previously in-house developed multimodality-guided synergistic network (MMgSN-Net) was applied to learn the mapping from CTP to CT Plain. The yielded Predicted CT Plain images were assessed against the ground-truth CT Plain using different imaging features, including blood vessel and lesions. Image quality was evaluated using spatial resolution and the overall authenticity. A quantitative analysis test rated the above components was designed (5 point visual scale, 5 = best, 1= worst).
Results: The qualitative evaluation showed that the Predicted CT Plain successfully removed the contrast enhancement from CTP, yielding a close visual approximation to the real CT Plain. This is indicated in the wash out of blood vessel enhancements and the consistence of tumor presentation. Spatial resolution of the predicted images is acceptable as the lesion margin is easily resolved. These assessments showed the Predicted CT Plain has high potential for clinical usage, achieving an average score of 3.65/5 for overall image appearance, 3.8/5 for CT image quality and 3.75/5 for the overall authenticity.
Conclusion: This study demonstrated that clinically equivalent synthetic CT Plain images can be generated from contrast-enhanced CT using the MMgSN-Net method, holding great promises for various applications.
Funding Support, Disclosures, and Conflict of Interest: GRF 15102219 MHP/005/20 Shenzhen-Hong Kong-Macau S&T Program (Category C) (SGDX20201103095002019) Shenzhen Basic Research Program (R2021A067)
Keywords
CT, Image Analysis, Radiography
Taxonomy
IM- CT: Machine learning, computer vision
Contact Email