Purpose: A strong prognostic score that enables a stratification of newly diagnosed Hodgkin Lymphoma (HL) to identify patients at high risk of refractory/relapsed disease is still needed. Our aim was to investigate the potential value of a radiomics analysis pipeline from whole-body MRI (WB-MRI) exams for clinical outcome prediction in patients with HL.
Materials and methods: Index lesions from baseline WB-MRIs of 40 patients (22 females; mean age 31.7 ± 11.4 years) with newly diagnosed HL treated by ABVD chemotherapy regimen were manually segmented on T1-weighted, STIR, and DWI images for texture analysis feature extraction. A machine learning approach based on the Extra Trees classifier and incorporating clinical variables, 18F-FDG-PET/CT-derived metabolic tumor volume, and WB-MRI radiomics features was tested using cross-validation to predict refractory/relapsed disease.
Results: Relapsed disease was observed in 10/40 patients (25%), two of whom died due to progression of disease and graft versus host disease, while eight reached the complete remission. In total, 1403 clinical and radiomics features were extracted, of which 11 clinical variables and 171 radiomics parameters from both original and filtered images were selected. The 3 best performing Extra Trees classifier models obtained an equivalent highest mean accuracy of 0.78 and standard deviation of 0.09, with a mean AUC of 0.82 and standard deviation of 0.08.
Conclusions: Our preliminary results demonstrate that a combined machine learning and texture analysis model to predict refractory/relapsed HL on WB-MRI exams is feasible and may help in the clinical outcome prediction in HL patients.
Keywords: Hodgkin Lymphoma; Machine learning; Magnetic resonance imaging; Positron emission tomography; Texture analysis.
Copyright © 2021 Elsevier Inc. All rights reserved.