Deep Learning-Based Prognosis Prediction of Diffuse Large B-Cell Lymphoma in PET Images
C Qian1;3*, C Jiang2, X Kai1, D Cao3, J Sun1, G Liugang1, N Xinye1, (1) The Affiliated Changzhou NO.2 People's Hospital of Nanjing Medical University, Changzhou, Jiangsu, CN, (2) Nanjing Drum Tower Hospital, Nanjing, Jiangsu, CN, (3) Changzhou Institute of Technology, Changzhou, Jiangsu, CN
Presentations
TU-D930-IePD-F2-5 (Tuesday, 7/12/2022) 9:30 AM - 10:00 AM [Eastern Time (GMT-4)]
Exhibit Hall | Forum 2
Purpose: Diffuse large B-cell lymphoma (DLBCL) is the most common form of non-Hodgkin lymphoma among adults. The accurate prognosis prediction of DLBCL in PET images is needed to guide new therapeutic strategies.
Methods: In this study, we propose a new framework for DLBCL prognosis prediction. Our framework consists of R-signature construction and prognosis prediction. Our R-signature construction includes four steps: deep learning (DL)-based feature extraction, first machine learning (ML)-based feature selection, second ML-based feature selection and R-signature generation based on classification/regression method. The DL-based features, metabolic metrics, and clinical risk factors are together used for prognosis prediction, especially progression-free survival (PFS) and overall survival (OS) predictions. Furthermore, we develop the survival prediction models using Cox regression analysis. Finally, the calibration and performance of the models are assessed and externally validated.
Results: In our work, a total of 398 PET images from patients with DLBCL are included. These images are from two different imaging centers. 288 images from the first imaging center constitute a training cohort and the validation cohort is composed of 110 images from the second imaging center. In the R-signature construction, 25 DL-based features are significantly associated with the prognosis prediction (P<0.05). In this study, the National Comprehensive Cancer Network International Prognostic Index (NCCN-IPI) is used for comparison. In the training cohort, our proposed method exhibited significant prognosis prediction superiority over NCCN-IPI in terms of both PFS (C-index: 0.762 vs. 0.697) and OS (0.757 vs. 0.696). In the validation cohort, the values are respectively 0.782 vs. 0.673 for PFS and 0.839 vs. 0.708 for OS. The calibration curves showed good consistency, and Kaplan-Meier analysis curves support the clinical utility of our proposed model.
Conclusion: The DL-based R-signature could be used as a prognosis predictor for DLBCL, and its combination with clinical factors may allow for accurate risk stratification.
Funding Support, Disclosures, and Conflict of Interest: We declare that we have no financial and personal relationships with other people or organizations that can inappropriately influence our work. This study is supported by General Program of Jiangsu Provincial Health Commission (No. M2020006), Natural Science Research Project of Colleges and Universities in Jiangsu Province (No. 21KJB110027).
Keywords
PET, Image Analysis, Feature Extraction
Taxonomy
IM- PET : Machine learning, computer vision
Contact Email