Utility of pre-treatment FDG PET/CT–derived machine learning models for outcome prediction in classical Hodgkin lymphoma

Hodgkin’s lymphoma (HL) is a haematopoietic malignancy characterised by the presence of Reed-Sternberg cells [1]. There are five different sub-classes of HL: nodular lymphocyte-predominant HL (NLPHL), and four under the umbrella category of classical HL (cHL): nodular sclerosing, mixed cellularity, lymphocyte-rich and lymphocyte-depleted. Ninety percent of HL cases are cHL [2]. NLPHL is often treated differently to cHL and is associated with more indolent progression [2]. Given the higher proportion of cHL cases, difference in treatment regimens and higher relapse rate in cHL compared to NLPHL, this paper will focus on cHL only [3].

Chemotherapy is the mainstay of frontline treatment of cHL; the most common regimes being doxorubicin (Adriamycin), bleomycin, vinblastine and dacarbazine (ABVD), or bleomycin, etoposide, doxorubicin (Adriamycin), cyclophosphamide, vincristine (Oncovin), procarbazine, and prednisone (BEACOPP) [4]. The treatment regime and number of cycles can vary depending on patient risk factors, disease stage and initial treatment response. Radiotherapy is used in patients with stage 1 or localised stage 2 disease or in residual bulky disease [4]. The gold standard imaging modality for staging and response assessment in HL is 2-deoxy-2-[¹⁸F]fluoro-d-glucose (FDG) positron emission tomography/computed tomography (PET/CT) [5]. Patients typically undergo PET/CT pre-treatment, following two cycles of chemotherapy (interim) and post-treatment. Interim PET/CT is used to guide treatment adaption, balancing the risk of chemotherapy-associated toxicity with maximising chances of event-free survival (EFS) [6]. Five-year survival in HL is approximately 86% [7]. However, even following complete metabolic response (CMR), approximately 20% of cHL patients will relapse with 72% of relapses occurring within the first 2 years of diagnosis [8]. The ability to identify patients at greater risk of relapse pre-treatment would allow upfront treatment stratification and could improve outcomes.

Previous studies assessing imaging parameters derived from baseline PET/CT for outcome prediction have mainly focused on metabolic tumour volume (MTV), total lesion glycolysis (TLG) and maximum or mean standardised uptake value (SUVmax and SUVmean) [9]. SUV is defined as the ratio of injected radioactivity within an image at a given timepoint when compared to the whole-body [10]. MTV is the volume of metabolically active segmented disease, with different segmentation techniques described [11]. The TLG is MTV multiplied by the SUVmean. Radiomics transforms images into mineable high-dimensional data permitting invisible feature extraction, analysis and modelling [12]. A limited number of studies using small sample sizes have demonstrated the potential of radiomic features in predicting progression-free survival (PFS) or overall survival (OS) in HL patients [13‐16]. The aim of this work was to evaluate the performance of models using radiomic features derived from pre-treatment FDG PET/CT to predict 2-year EFS in cHL patients using a larger tertiary centre cohort of patients.

This study adhered to the transparent reporting of a multivariable prediction model for individual prognosis or diagnosis (TRIPOD) guidelines (Supplemental Material 1).

This study confirms that pre-treatment outcome prediction using FDG PET/CT–derived radiomic features is feasible in patients with cHL. The best performing model was created using ridge regression combining age and four radiomic features (PET flatness, PET major axis length, PET logarithm GLSZM size zone non-uniformity normalised, PET lbp-3D-m1 GLCM correlation and PET lbp-3D-m2 first order skewness) extracted from PET images using a 1.5 × mean liver SUV method with a bin width of 0.24. It must be noted that there was no significant difference between the test AUC of this model and those of a combined clinical and MTV model and a model created using 4.0 SUV fixed threshold segmentation. This is likely due to small numbers involved given the relatively large confidence intervals. Due to missing clinical data, it was not possible to adjust for features used to stratify patients into early and advanced disease. A surrogate, treatment intent, was used instead which demonstrated that the models created remained reasonable predictors of outcome for patients treated as having advanced disease.

Further work should be performed to assess the relationship of ethnicity and socio-economic status on a model’s predictive ability to avoid creation of a model which discriminates against under-represented subsets of patients due to lack of data to train and test the model on [21, 22]. Unadjusted confounders are likely one of the reasons for a minority of studies reporting the poor ability of MTV as an outcome predictor in lymphoma [23‐25]. Most notably Adams et al found that MTV was not an independent predictor of overall survival or PFS in diffuse large B cell lymphoma once adjusting for the National Comprehensive Cancer Network International Prognostic Index [26]. To allow for transparency, our study has provided the demographic information for the patients who were mislabelled using the predictive model with the highest test AUC.

Two different segmentation techniques were explored. The first was a fixed threshold of 4.0 SUV which has been demonstrated to be a reproducible, efficient method for contouring disease [27]. The second was 1.5 × mean liver SUV which has been explored in other malignancies and provides an adaptive threshold which adjusts for background SUV uptake [17, 18]. Our study echoed previous work demonstrating a fixed threshold led to more features being robust following re-segmentation. The fixed thresholding segmentation technique required less steps, and less manual adaption [28]. However, a fixed SUV thresholding technique does not scale with the physiological uptake and therefore, the contours may vary on repeat studies due to external effects on the SUV rather than tumour pathophysiology [27]. The study also demonstrated the variability which can occur when repeating a segmentation methodology on different software (Fig. 5), with radiomic features not being deemed robust following repeated segmentation even when using the same SUV thresholds. ComBat harmonisation was employed to mitigate against the effects of scanner variation. This is based on Bayes theorem and attempts to predict scanner influence whilst maintaining biological variation [29]. For this to be effective, however, there must be enough samples from different scanners to apply the harmonisation method [30] and it cannot be applied prospectively to scanner acquisitions outside those used for training of the predictive model.

Previous studies have explored the use of radiomic features in the prediction of outcomes in HL [13‐15]. Lue et al found SUV kurtosis, stage and intensity non-uniformity (INU) derived from Grey Level Run Length Matrix (GLRLM) were independent predictors of PFS in a small cohort of 42 patients. Milogrom et al demonstrated that the combination of SUVmax, MTV, InformationMeasureCorr1, InformationMeasureCorr2 and InverseVariance derived from GLCM 2.5 had an AUC of 0.95 when predicting relapse in 167 patients with stage I–II HL. However, there were very few events, with the validation cohort only having two patients who relapsed. Sollini et al assessed a radiomic fingerprint using principal component analysis to classify patients who would relapse within 4 years of treatment in a cohort of 85 patients. They explored fingerprints created from a single largest nodal or extra-nodal lesion versus using all lesions and found that the intra-patient similarity was low, and that the highest accuracy was achieved when using all lesions within the model [15]. This highlights the inherent heterogeneity of radiomic features within different lesions and that by restricting analysis to a single lesion, the predictive model may also be limited. The current study of 289 patients is one of the largest to assess potential utility of radiomic features derived from pre-treatment FDG PET/CT for predicting outcome in cHL patients. It demonstrates that radiomics could feasibly improve prediction of 2-EFS. However, this requires validation on an independent external dataset and although the AUC for the test set was 0.81, no clear predictive threshold could be derived. This must be a key target when creating any machine learning or AI-based model. In terms of HL, it would be the ability to balance side effects of escalated treatment, with the rates of EFS and toxicity varying between treatment regimens [31]. The advent of newer therapeutic strategies limits the use of predictive models made on retrospective data; future efforts should focus on validating imaging, genetic and clinical predictive features in carefully designed prospective, multi-centre clinical trials.

A TRIPOD checklist was used to ensure transparency of the study’s methodology, a concern in previous radiomic studies [32, 33]. However, no external validation was performed, and although contouring was undertaken without knowledge of clinical outcome, no measures to blind assessors were specifically undertaken. Although patients with other concurrent malignancies were excluded from analysis, other pathologies were not taken into consideration when looking at mortality. Other study limitations include its retrospective nature, the relatively small event rate, reliance on clinical records to determine date of relapse/recurrence, exclusion of patients with hepatic disease/or without disease > 4.0 SUV and variation in different patient’s treatment regimen.

There is potential for models derived from radiomic features extracted from pre-treatment FDG PET/CT to predict 2-EFS in cHL patients. Further work is needed to determine optimum thresholds for clinical use.