Distinguishing diffuse large B-cell lymphoma from Hodgkin's lymphoma in children using an enhanced computed tomography radiomics approach.

Abstract

Diffuse large B-cell lymphoma (DLBCL) and Hodgkin's lymphoma (HL) are two completely different pathologic subtypes of lymphoma with distinctly different clinical presentations and treatment options. Thus, accurately differentiating between the two subtypes has important clinical implications. This study aimed to construct a radiomics model capable of distinguishing between DLBCL and HL based on enhanced computed tomography (CT) for the non-invasive diagnosis of lymphoma subtypes. The clinical and imaging data of 16 patients confirmed to have DLBCL (33 lymphomas), and 50 patients confirmed to have HL (106 lymphomas) were retrospectively analyzed. The patients were completely randomized into a training set (n=107, DLBLC׃HL ratio: 23׃84) and a test set (n=32, DLBCL׃HL ratio: 10׃22). After multiple down-sampling, 2,264 radiomics features were automatically extracted by the application software. Feature selection was performed in the training set using Spearman's rank correlation coefficients, maximum correlation minimum redundancy, and the least absolute shrinkage and selection operator algorithm in that order. The features after selection were used to build radiomics models by logistic regression (LR) and quadratic discriminant analysis (QDA). We evaluated the model ability using receiver operating characteristic (ROC) curves and the DeLong test. Moreover, clinical indicators, such as gender, age, clinical stage, and lactate dehydrogenase (LDH), were collected and analyzed by univariate and multivariate LR analyses. The radiomics characteristics with clinical indicators that had independent influences on predicting the pathological subtypes were used to establish a comprehensive classification model. The analysis of the clinical data revealed that LDH can serve as a clinical indicator that has an independent influence on the prediction of HL and DLBCL. The results of the radiomics models were as follows: Radiomics_LR: area under the curve (AUC) =0.814 [95% confidence interval (CI): 0.628-0.999]; and Radiomics_QDA: AUC =0.841 (95% CI: 0.691-0.991). Following the inclusion of LDH as a clinical indicator in the analysis, the results of the comprehensive models were as follows: Radiomics + LDH_LR: AUC =0.768 (95% CI: 0.580-0.956); and Radiomics + LDH_QDA: AUC was 0.845 (95% CI: 0.695-0.996). The models based on radiomics and clinical features were able to effectively distinguish DLBCL from HL. The model with the best overall performance was the Radiomics_LR model.