Objective To assess whether a diagnostic analysis model based on radiomic features extracted from clinical baseline ¹⁸F-fluorodeoxyglucose (FDG) PET is useful in differentiating Hodgkin lymphoma (HL) from aggressive non-Hodgkin lymphoma (NHL).

Methods Clinical data and ¹⁸F-FDG PET imaging data of 182 patients with lymphoma confirmed by histopathological results, who were treated at the First Affiliated Hospital of Zhengzhou University from May 2016 to July 2021, were retrospectively analyzed, including 111 males and 71 females, age 4−85 years old with a median age of 50. All patients were divided into a training group (110 patients) and a validation group (72 patients) in a 6∶4 ratio by simple random sampling. The radiomic features in the ¹⁸F-FDG PET images were extracted using 3D Slicer, and the extracted radiomic features of the training groups were subjected to an independent sample t-test, from which statistically significant differences in the radiomic features were screened out. The radiomic features obtained from the initial screening were screened again using least absolute shrinkage and selection operator (LASSO) regression in the training group to obtain the best subset of radiomic parameters. Predictive models were constructed using multivariate Logistic regression analysis. The efficacy of the diagnostic model for differential diagnosis of HL and NHL by radiomic features was evaluated using the receiver operating characteristic (ROC) curve. The Mann-Whitney U rank-sum test was used for comparisons of measurements that did not conform to normal distribution, whereas χ² test was used for comparisons of count data.

Results Among the 110 patients in the training group, 33 had HL and 77 had NHL. Among the 72 patients in the validation group, 22 had HL and 50 had NHL. The differences in SUV_max, metabolic tumor volume, total lesion glycolysis between the patients with HL and NHL in the training and validation groups were not statistically significant (Z=0.232−1.689, all P>0.05). The optimal λ-value of the LASSO regression was 0.030, and in accordance with the optimal λ-value of the radiomic parameters, 13 radiomic features were obtained, and 8 radiomic features were screened out by multivariate Logistic regression analysis. The area under curve (AUC) of ROC of the radiomic feature model in the training group was 0.880 (95%CI: 0.802−0.958), and the sensitivity, specificity, and accuracy were 83.5%, 65.5%, and 78.2%. The AUC of ROC of the radiomic feature model in the validation group was 0.706 (95%CI: 0.570−0.842), with sensitivity, specificity, and accuracy of 70.9%, 81.8%, and 74.2%, respectively.

Conclusion Models based on PET radiomic features has great potential application value in the differential diagnosis of HL and invasive NHL.