The role of 18F-FDG PET/CT in the prediction of Hodgkin lymphoma therapy assessment as well as Deauville score

1407 Introduction: Treatment management decisions for Hodgkin lymphoma (HL) patients are based on [18F]FDG-PET/CT and hioptic data. To date, the association of PET and CT-derived in vivo radiomic features with HL subtype and Deauville score prediction is not well explored. The aim of this study was to investigate the feasibility of predicting therapy ressponce and Deauville score with the help of machine learning.Materials and Methods: 118 [18F]FDG-PET/CT patients were included in this study. All cases had documented demographics (age, gender, weight, height and BMI) and clinical information (bone marrow, spleen and lung involvement, stadium according to Lugano classification, systemic symptoms, chemo and/or radio-therapy, number of cycles and protocol employed). The delineation on PET/CT was PET-driven (Hermes Nuclear Diagnostics, Sweden) followed by optimized radiomics feature extraction. Afterwards, therapy response (TR) 1-vs-2/3, 1/2-vs-3 (1-metabolic regression, 2-partial regression and 3-progression) and Deauville score (1/2-vs-3/4/5) predictor models were established with ensemble learning approach. Monte Carlo cross-validation with 90% training and 10% validation sets in 100-fold cross-validation scheme was performed to estimate the sensitivity (SENS), specificity (SPEC), accuracy (ACC), positive-predictive-value (PPV) and negative-predictive-value (NPV) of all predictive models. Class imbalance was handled with random undersampling. Results: All predictive models performed moderately well (Table 1). The Deauville as well as the TR 1-vs-2/3 predictive models had moderate SPEC (62% and 60% respectively), while the TR 1/2-vs-3 predictive model had a SENS of 61%. The most important feature categories in the three predictive models were: Deauville (CT, 42%), TR 1-vs-2/3 (CT, 42%), TR 1/2-vs-3 (CT, 42%). See Table 2 for the PET, CT, PET+CT, demographics and clinical feature weights of all predictive models. Conclusions: The preliminary results of our work imply that more and homogeneous data is required preferably from multi-centric environments. Such data is already available and under evaluation as part of this study.