Introduction
Diffuse large B-cell lymphoma (DLBCL) represents the most common type of lymphoid neoplasm [
1]. Since over 30% of patients experience disease progression or relapse, early identification of high-risk patients is important for patient management [
2]. Over the past two decades, the International Prognostic Index (IPI) has been recognized as a prognostic model, which is based on the properties of several clinical factors including age, Ann Arbor stage, extranodal involvement, serum lactate dehydrogenase (LDH) level, and performance status [
3]. However, IPI is not suitable for predicting refractory disease, which might be due to its lack of information on intratumoral functional and metabolic profiles [
4,
5].
Positron emission tomography/computed tomography (PET/CT) with 2-deoxy-2-[
18F]fluoro-D-glucose ([
18F]FDG) is a representative of molecular imaging and transpathology [
6], which has been applied as a routine imaging tool for staging and response assessment of lymphoma [
7]. Several studies have indicated that PET semi-quantitative parameters, particularly maximum standardized uptake value (SUVmax), total metabolic tumor volume (TMTV), and total lesion glycolysis (TLG), might be independent prognostic factors in DLBCL [
8‐
10]. However, those parameters are only used to evaluate the gross tumor metabolism, which cannot fully depict the subtle metabolic heterogeneity within a targeted lesion. Recently, PET-based radiomics has been introduced as an innovative image analysis that can capture intratumoral metabolic heterogeneity and allow accurate prediction of clinical outcome in various malignancies, such as breast cancer, non-small cell lung cancer, and lymphoma [
11‐
13]. Studies have shown that several single radiomic features, including long-zone high gray-level emphasis, and skeletal textural feature SkewnessH, were significant predictors of survival in DLBCL [
14,
15]. Other literatures reported that the combination of multiple radiomic features, which is often defined as the radiomic signature (RS), may hold higher prognostic value than the single feature [
16,
17]. In addition, RS combined with clinical or genomic data can produce robust and improved medical decision-making [
18,
19]. However, to the best of our knowledge, the RS based on [
18F]FDG PET/CT for prognosis assessment of DLBCL has not yet been described. Furthermore, it remains unclear whether PET-based RS could add more prognostic values to the IPI in DLBCL.
We hypothesized that RS combined with IPI score could help improve the prognosis assessment of DLBCL patients. Therefore, for the first time, this study investigated the prognostic value of the RS combined with IPI score in predicting the survival of DLBCL patients by [18F]FDG PET/CT.
Discussion
In the present study, we developed RS and hybrid nomograms from pre-treatment [18F]FDG PET/CT images for outcome prediction in patients with DLBCL. The RS and IPI were identified as independent predictors of PFS and OS. The hybrid nomograms combining RS with IPI performed better than IPI alone, indicating that the RS could provide incremental prognostic values to the IPI. To the best of our knowledge, this is the first study that combines RS with IPI to assess the intratumoral heterogeneity on pre-treatment [18F]FDG PET/CT in DLBCL.
The most important finding of our study was that the RS composed of multiple radiomic features could improve the prognostic value beyond the conventional IPI score. The hybrid nomograms combining RS with IPI could help stratify those high-risk individuals with poorer survival outcomes, achieving significantly higher AUCs and contributing to more distinct risk stratifications than IPI alone. This is consistent with a very recent study indicating that a single radiomic feature run length non-uniformity could provide additional prognostic value to the IPI in DLBCL [
35]. However, compared with the run length non-uniformity reported in their study, the RS in our study showed more significant
P values in the multivariate analysis (e.g., the
P value of the MBV-based RS for PFS prediction was 0.001). Similarly, another study also identified a single radiomic feature long-zone high gray-level emphasis as an independent predictor of 2-year event-free survival (with a sensitivity of 0.60) [
14]. By comparison, the RS in our study showed higher sensitivities for survival prediction (e.g., the MBV-based RS had a sensitivity of 0.84 for PFS prediction in the training cohort). In our study, neither MBV- nor TMTV-based RS showed significantly higher AUCs than IPI score. This observation is in line with previous studies demonstrating that the predictive ability was comparable between RS and clinical variables [
36,
37]. A possible explanation is that IPI score and RS possess different properties in phenotyping disease characteristics. IPI score is based only on clinical factors, while RS represents imaging features that reflect the intratumoral metabolic heterogeneity. Since the complex nature and biologic processes of malignancy involve multiple components, taking both clinical and imaging features into account may provide a more comprehensive disease characterization and a better prognostication. In this study, we reported the first attempt to develop hybrid nomograms by combining the PET-based RS with IPI, which was supposed to provide an individualized estimate of survival and could serve as an easy-to-use tool for clinical decision-making. Taken previous findings and our results together, we speculated that PET-based radiomics and IPI could be complementary and synergistic for estimating survival in DLBCL.
Radiomic analysis for lymphoma is challenging, at least in part due to the inter- and intratumoral heterogeneity, and the complexity of isolated lesion segmentation especially when disease is disseminated [
38]. In light of these concerns, we performed radiomic analysis on the metabolic volume of the largest lesion, which was defined as MBV in our study, and compared its performance with that based on TMTV. Our results demonstrated that radiomic analysis on MBV and TMTV both perform well in predicting survival, which is in line with previous reports [
13,
14]. Moreover, no significant difference was found between the performance of MBV- and TMTV-based hybrid nomograms. As measuring MBV is technically easier and faster than measuring TMTV, our results indicated that radiomic analysis on MBV could be a feasible approach for prognosis assessment in DLBCL.
We also compared the prognostic value of MBV and TMTV. While TMTV has been commonly reported as a potential prognostic indicator in DLBCL [
9,
39], very few studies have focused on the prognostic effect of MBV. A recent study suggested that MBV was an independent predictor of OS and had a strong correlation with TMTV [
24]. Consistently, ROC analysis in our study showed no significant difference between MBV and TMTV in survival prediction, suggesting that MBV holds prognostic value as TMTV does. Besides, for 25% of patients in validation cohort who had discordance in MBV and TMTV, MBV accurately predicted the outcome regardless of TMTV, which was in accordance with the previous finding that MBV might have greater influence on survival than TMTV [
24], and indicated that MBV could be a surrogate marker of TMTV.
Our results showed that conventional PET parameters including SUVmax, TMTV, and TLG were not retained for model construction. However, these parameters were reported to be predictive of survival in DLBCL [
8‐
10]. This discrepancy may be attributed to the differences in the methods employed for feature selection and model building. In our study, we applied the LASSO Cox regression algorithm which is considered suitable for screening high-dimensional features that are most strongly associated with patient outcome and avoiding overfitting [
40]. As shown in our results, only radiomic features were finally selected via this algorithm, indicating that radiomic features were correlated with tumor volume and might provide more accurate prognostic information than conventional PET parameters in DLBCL.
In our study, we applied the 41% SUVmax method which has been recommended by the EANM and identified as an effective approach for prognosis assessment in DLBCL [
21,
41,
42]. The results of the ICC analysis showed that the majority of assessed features had good intra- and inter-observer agreement (ICC > 0.75), which is consistent with a recent study demonstrating that the 40% SUVmax method could improve the repeatability of most radiomic features [
43]. However, it has been revealed that the radiomic features could be influenced by different segmentation methods [
44]. Future studies are required to use multiple segmentation and explore optimized methods through more advanced, deep learning techniques [
45,
46].
There are some limitations to this study. First, protein expressions and gene arrangements of MYC and BCL-2 are acknowledged prognostic factors but are not evaluated in our study due to the unavailability of these data from all patients. Second, one should be cautious when extrapolating these findings as this is a retrospective single-center study with a relatively small sample size. Therefore, our results need to be further validated in prospective multi-center studies involving a larger cohort of patients.
Publisher's Note
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.