Abstract Introduction Peripheral T-cell lymphoma (PTCL) is a heterogeneous group of aggressive lymphoma. While novel therapies are emerging, molecular and histopathological analyses for precise classification remain limited in clinical practice, and prognostic prediction with the International Prognostic Index (IPI), or the Prognostic Index for PTCL-U (PIT) is suboptimal. Radiomic features (RFs) may offer a promising approach to quantify tumor heterogeneity at the voxel level, potentially reflecting underlying metabolic patterns. This study evaluates RFs in PTCL for improved diagnostic differentiation and prognostic stratification. Methods This retrospective study included patients histopathologically diagnosed with nodal T-follicular helper (TFH) lymphoma or PTCL-not otherwise specified (PTCL-NOS) at Kameda Medical Center between 2007 and 2024. Pathologic lesions on positron emission tomography-computed tomography (PET-CT) were delineated using a 41% threshold of maximum standardized uptake value (SUVmax). RFs were extracted from the hottest lesion with volume 3 mL using PyRadiomics. All variables were normalized prior to model development. Diagnostic models were developed using the least absolute shrinkage and selection operator (LASSO) logistic regression to differentiate TFH and PTCL-NOS, while prognostic models used LASSO Cox regression to predict progression-free survival (PFS), both validated with 10-fold cross-validation. For each task, two model types were created: one incorporating RFs and one without, alongside clinical parameters from IPI, PIT, and conventional PET metrics, including SUVmax and total lesion glycolysis (TLG). Discriminative performance was evaluated using receiver operating characteristic (ROC) analysis and DeLong's test. Decision curve analysis (DCA) compared model utility. Prognostic stratification was assessed using Kaplan-Meier analysis with log-rank testing, and predictive accuracy was evaluated using Harrell's C-index. Results A total of 108 patients were included, comprising 56 (48. 7%) with TFH and 52 (45. 2%) with PTCL-NOS. The median age was 74. 5 years (interquartile range: 68. 0–81. 0), with 40 (37. 0%) females and 68 (63. 0%) males. All patients received anthracycline-based chemotherapy; 9 (8. 3%) underwent upfront autologous stem cell transplantation. In diagnostic prediction, five RFs were selected for the model incorporating radiomics, including increased asymmetry in voxel intensity patterns (gray level co-occurrence GLCMCluster Shade; coefficient: 0. 42) favoring TFH, and uniformity in intensity distribution (GLCMMaximal Correlation Coefficient; -0. 39) favoring PTCL-NOS. Additional features included number of extranodal lesions (EN; -0. 77) and SUVmax (-0. 67). This model achieved an AUC of 0. 902, outperforming the model without radiomics (AUC = 0. 807; p = 0. 002). The radiomics-based model also showed greater clinical utility across all risk thresholds, as assessed by DCA. For prognostic modeling, two RFs were selected for the radiomics-based model, including small, low-avidity lesions (gray level size zone matrixSmall Area Low Gray Level Emphasis; coefficient: 0. 12), alongside EN (0. 39) and TLG (0. 33). When patients were stratified into three risk categories, the model significantly differentiated PFS (median: 56. 3, 23. 6, 4. 9 months; p 0. 001) and overall survival (OS; 103. 6, 57. 0, 8. 9 months; p 0. 001). It also demonstrated the highest predictive accuracy for PFS (C-index = 0. 68) and comparable performance for OS (0. 71), outperforming the model without RFs (0. 66 and 0. 70), IPI (0. 62 and 0. 68), and PIT (0. 61 and 0. 66). In subgroup analysis, only the radiomics-based model significantly stratified PFS in TFH (median: 56. 3, 24. 2, 9. 3 months; p = 0. 001), while the non-radiomics model did not (p = 0. 12). In PTCL-NOS patients, all four models significantly stratified both OS and PFS. Conclusion This study demonstrates the diagnostic and prognostic utility of radiomic features in PTCL, highlighting their potential to enhance risk stratification beyond conventional PET metrics and clinical indices. Further research incorporating molecular and histopathological subclassification is warranted to validate these findings and clarify the biological significance of the extracted features.
Sakuma et al. (Mon,) studied this question.