PET/CT radiomics-biomarker diagnostic model for identifying lymphoma in fever of unknown origin.
Fever of unknown origin (FUO) with lymphadenopathy poses a substantial diagnostic challenge, particularly in differentiating malignant lymphoma from benign lymph node disorders. Although PET/CT plays a crucial role in lesion detection, inflammatory and immune-mediated lymphadenopathies often exhibit metabolic patterns overlapping with lymphoma, leading to limited diagnostic specificity. Radiomics enables high-throughput extraction of quantitative imaging features that capture intratumoral heterogeneity beyond visual assessment, while blood-based biomarkers reflect systemic inflammatory and metabolic states. This study aimed to develop and validate a multimodal diagnostic model integrating PET/CT radiomics and clinical biomarkers to improve the discrimination of lymphoma from benign lymphadenopathy in patients with FUO.
This retrospective study included FUO who underwent PET/CT. Lymph nodes were selected if the short-axis diameter was ≥1 cm on CT or if metabolic activity exceeded the mediastinal blood pool. Clinical and laboratory data were collected, and radiomics features were extracted from 40%SUVmax-based VOIs using LIFEx. Patients were randomly divided into training and testing sets (7:3). A biomarker model, radiomics model, and combined model were constructed using logistic regression, with feature selection performed via the least absolute shrinkage and selection operator (LASSO) for radiomics. Model performance was evaluated using receiver operating characteristic (ROC) curves, and DeLong's test. A nomogram was developed for individualized prediction.
A total of 203 patients were enrolled (114 lymphoma, 89 benign). In the training cohort, under the curve (AUC)s for the biomarker, radiomics, and combined models were 0.924, 0.903, and 0.970, respectively, with the combined model showing significantly superior performance. In the testing cohort, the corresponding AUCs were 0.903, 0.912, and 0.958, again demonstrating the highest accuracy for the combined model. Key predictors in the final model included procalcitonin (PCT), serum amyloid A (SAA), and PET/CT radiomics features. A nomogram was generated to facilitate individualized risk estimation.
PET/CT radiomics provides strong discriminatory ability for differentiating lymphoma from benign lymphadenopathy in FUO. Incorporating multidimensional biomarkers such as PCT and SAA further enhances diagnostic performance.
This retrospective study included FUO who underwent PET/CT. Lymph nodes were selected if the short-axis diameter was ≥1 cm on CT or if metabolic activity exceeded the mediastinal blood pool. Clinical and laboratory data were collected, and radiomics features were extracted from 40%SUVmax-based VOIs using LIFEx. Patients were randomly divided into training and testing sets (7:3). A biomarker model, radiomics model, and combined model were constructed using logistic regression, with feature selection performed via the least absolute shrinkage and selection operator (LASSO) for radiomics. Model performance was evaluated using receiver operating characteristic (ROC) curves, and DeLong's test. A nomogram was developed for individualized prediction.
A total of 203 patients were enrolled (114 lymphoma, 89 benign). In the training cohort, under the curve (AUC)s for the biomarker, radiomics, and combined models were 0.924, 0.903, and 0.970, respectively, with the combined model showing significantly superior performance. In the testing cohort, the corresponding AUCs were 0.903, 0.912, and 0.958, again demonstrating the highest accuracy for the combined model. Key predictors in the final model included procalcitonin (PCT), serum amyloid A (SAA), and PET/CT radiomics features. A nomogram was generated to facilitate individualized risk estimation.
PET/CT radiomics provides strong discriminatory ability for differentiating lymphoma from benign lymphadenopathy in FUO. Incorporating multidimensional biomarkers such as PCT and SAA further enhances diagnostic performance.
Authors
Jing Jing, Zhang Zhang, Hua Hua, Zhang Zhang, Li Li, Cai Cai, Tian Tian, Wei Wei, Bian Bian
View on Pubmed