Radiomics-Based Multiple Prognosis Prediction Using Mutual Information Between Treatment Outcomes as Prior Knowledge in Nasopharyngeal Carcinoma
Z Ma1*, J Zhang1, X Teng1, S Lam1, X Han1, H Xiao1, C Liu1, W Li1, Y Huang1, F Lee2, W Yip3, A Cheung1, H Lee4, J Cai1, (1) The Hong Kong Polytechnic University, Hong Kong (2) Queen Elizabeth Hospital, Hong Kong (3)Hong Kong Sanatorium and Hospital(4) The University Of Hong Kong
Presentations
SU-H400-IePD-F5-2 (Sunday, 7/10/2022) 4:00 PM - 4:30 PM [Eastern Time (GMT-4)]
Exhibit Hall | Forum 5
Purpose: We hypothesize that multiple prognosis prediction incorporating correlations among treatment failure outcomes could lead to better performance than multiple separate predictions for each outcome. This study aimed to develop a radiomics-based multiple prognosis prediction model that incorporates mutual information of each treatment failure in nasopharyngeal carcinoma (NPC).
Methods: We enrolled 221 patients with stage III-IVA NPC patients who were treated by concurrent chemoradiotherapy from December 2012 to December 2016. All the patients had at least three years of follow-up time. Four treatment outcomes, including Regional Recurrence (RR), Distant Metastasis (DM), Local Recurrence (LR) and Progression-Free Survival (PFS), were predicted using binary relevance (BR) and classifier chain (CC) models. Patients were randomly split into training and testing sets (7:3 ratio for 15 times). A total of 5580 radiomics features were extracted from contrast-enhanced computed tomography (CECT). Radiomics features were selected by using the least absolute shrinkage and selection operator, and models were constructed using the support vector machine with the linear kernel. For the BR method, each single outcome was trained independently. For the ECC method, 24 different CC models with random classifier orders were trained. The averaged accuracy was adopted to evaluate and compare the model performance.
Results: The average accuracy for the testing set was 0.677 and 0.715 in BR and CC models, respectively. The CC model with order RR > LR > PFS > DM showed highest testing accuracy of 0.824 (95% confidence interval, 0.805-0.845). The CC model showed higher accuracy than the BR model on the 4-prognosis outcomes (P < 0.1).
Conclusion: We built a radiomics-based multiple prognosis prediction chained model that incorporates mutual information of each treatment outcome in NPC. The chained model performed better than a simple bundle of binary classifiers incorporating outcomes’ relationships.
Funding Support, Disclosures, and Conflict of Interest: ITS/080/19 Project of Strategic Importance(P0035421) The Hong Kong Polytechnic University, and Shenzhen-Hong Kong-Macau S&T Program (Category C) (SGDX20201103095002019) Shenzhen Basic Research Program (R2021A067)
Keywords
Quantitative Imaging, Mutual Information, Modeling
Taxonomy
Contact Email