In our study, we evaluated the utility of combined PET/CT radiomics features as well as clinical parameters for outcome prediction of patients with Hodgkin’s lymphoma (HL).
So far, only a few radiomics studies have been performed in HL populations addressing outcome prediction and even fewer have considered clinical parameters as well as combined PET and CT features for outcome prediction. We found that CT as well as PET radiomics combined with clinical parameters might be able to help predict outcome endpoints like PFS and OS as well as the need for additional radiotherapy.
We found several radiomics parameters from baseline FDG PET/CT to be predictors of survival and predictors for the need for radiotherapy in the univariable analysis. However, when multivariable models were designed, considering the parameter with the lowest p-value for the model building, no PET-related parameter was found to be an independent predictor for the composite survival. This is concordant with a few earlier studies that evaluated a similar question, including first order parameters such as SUVmax. For example, Frood et al  recently published a meta-analysis for baseline PET/CT imaging parameters as predictor of treatment outcome in Hodgkin and diffuse large B-cell lymphomas (DLBCL). In the meta-analysis 10 studies assessing SUVmax as predictor of response, however, none none of the studies evaluated radiomics features. The largest study, by Akharti et al.  demonstrated that SUVmax could not be applied to predict either PFS or OS in 267 patients. Interestingly, in our study, a CT second order parameter was a predictor of survival when combined with clinical parameters such as albumin and ALP.
Driessen, J et al.  recently presented a radiomics analysis in a larger cohort of patients with relapsed HL. They found that a combination of radiomics and clinical features results in a strong prediction model for 3-year time to progression. The model uses robust PET features that address inter-lesional heterogeneity in distance, metabolic volume and SUV, but did not include any second or higher order radiomic features, as compared to our study. In addition, this investigation did not include radiomics evaluation of the CT- component of the PET/CT.
A recent study by Zhou and co-workers evaluated if the radiomic features of baseline FDG PET could predict the prognosis of Hodgkin lymphoma . They found that Long-zone high gray-level emphasis and Dmax were independently correlated with 2-year progression-free survival. Again, however, this study did not evaluate the complementary CT-radiomics, and did not integrate any clinical information into their AUC analysis. Furthermore, they evaluated a smaller number of patients, which were further divided into a training and validation data set which likely decreases statistical robustness.
Another study has taken a somewhat different approach, evaluating 45 patients receiving R-CHOP (Rituximab+ Cyclophosphamide + Doxorubicin + Vincristine + Prednisone) chemotherapy for DLBCL, evaluating the ability to predict therapy response . Here, the authors concluded that that SUVmax and gray level co-occurrence matrix dissimilarity were independent predictors of lesions with incomplete response.
Milgrom, S.A. et al.  analyzed a cohort of 251 mediastinal HL patents using another freely available software (IBEX). They found that first order parameters MTV and TLG are associated with disease progression in HL. None of the second order parameters were predictors of progression in their cohort either.
Lue et al.  investigated 11 first order, 39 higher order features in 42 patients with HL to predict PFS and OS. With 21 events in the cohort (12 relapses, 9 deaths) it was demonstrated that SUV, kurtosis, stage and intensity non-uniformity (INU) derived from Grey-Level Run Length Matrix (GLRLM) were independent predictors of PFS and only disease stage and INU derived from GLRLM were independent predictors of OS.
Overall, compared to the relatively sparse, directly comparable literature, in our study none of the PET-derived radiomic features were found to be independent features in the MVA for composite survival prediction. Since several PET-radiomic features were found to be significant in the UVA, if we had evaluated only PET radiomics features, it might be that those parameters would have been significant in the MVA and therefore, we would have more comparable results to other studies. However, in our investigation, PET radiomics parameters were ‘outperformed’ by the CT-radiomic features (which consequently ended up in the MVA) and were therefore not directly compared to the available studies. We feel that, since PET/CT is a hybrid imaging modality in clinical routine, both components (the PET and the CT) should be evaluated in a complementary fashion and as demonstrated, there appears to be value in CT-derived textural features as well.
However, in our cohort, a PET first order parameter, SHAPE Sphericity, and the CT second order features, GLZLM SZHGE mean and PARAMS XSpatial Resampling, were independent predictors for the need of radiotherapy when combined with hemoglobin result at baseline lab work (AUC=0.79) which again underlines the values for combined radiomic evaluation of PET and CT. It has to be pointed out that the clinical decision to apply additional radiotherapy is often multifactorial and that not only one clinical scenario indicates the need for radiotherapy in HL patients. However, based on our analysis, the integration of combined CT and PET radiomics features might be of further guidance/help to decide which patients might profit from additional radiotherapy for improvement of their disease outcome.
Again, similar to other studies cited above, only first order and morphologic PET radiomic features were found to be significant and thus, not necessarily intrinsically related with voxel characteristics. For CT, however, two second order features were found to be of value (i.e. GLZLM SZHGE). As for the comparable literature, no other studies evaluated predictors for the need for radiotherapy, besides the bulkiness of the tumor and therefore this finding may open a window for further analysis in larger cohorts.
Several other new studies have evaluated different aspects of radiomics i.e. in MRI or in PET, but those studies concentrated on technical aspects of the analysis itself rather than the ability of radiomics for prediction. Also, PET/CT radiomics has been thoroughly evaluated in non-Hodgkin lymphoma and in the context of prediction for bone marrow involvement, but as this was done for follicular lymphoma these studies are not specifically relevant for HL patients [41–44].
Concerning the integration of clinical parameters, it has been shown in the literature that ALP is not necessarily a predictive clinical parameter on its own. While that is certainly valid from a dedicated clinical perspective, in our cohort it has been found to be of predictive value in conjunction with the here evaluated imaging features. Thus, the integration of combined PET and CT radiomics features may elevate the value of specific clinical parameters when evaluated in conjunction.