Automated radiomics model for prediction of therapy response and minimal residual disease from baseline MRI in multiple myeloma

Abstract

Abstract This multicenter imaging study aimed to establish and validate automated radiomics models predicting therapy response (TR) and minimal residual response (MRD) in newly diagnosed multiple myeloma (MM) from baseline MRI. Retrospectively, 118 MM patients from the GMMG-HD7 trial (EudraCT: 2017-004768-37) with data on TR and/or MRD after induction therapy and baseline MRI were included. Data were split by center into a training set (center 1–2; n = 79) and a test set (center 3–10; n = 39). TR was classified as very good partial response or better versus other. An in-house developed nnU-Net was used to automatically segment pelvic bone marrow for the subsequent extraction of 245 radiomics features and piriformis muscle for normalization. Random forest classifiers were trained using radiomics features only (I), radiomics features with additional confounders (II) or myeloma-relevant clinical features (III), or only clinical features (IV) to predict TR or MRD status. The area under the receiver operating characteristic curve (AUROC) was calculated to assess prediction performance. The prediction model using only radiomics features (I) showed the highest predictive performance for TR on the test set with an AUROC of 0.70. AUROC values for radiomics-based prediction of the MRD status (I-III) ranged from 0.54 to 0.52. In conclusion, our study demonstrated the potential of automated radiomics models from baseline MRI to non-invasively predict TR in MM on an independent, multicentric test set.