Radiomic features of glucose metabolism enable prediction of outcome in mantle cell lymphoma

Mantle cell lymphoma (MCL) is a rare subtype of B cell non-Hodgkin lymphoma, and can be associated with an aggressive or, less frequently, an indolent course [1]. Despite the availability of novel types of treatment, the prognosis in MCL patients is generally considered to be poor [2], with 5-year survival rates as low as 50% [1]. For estimation of prognosis, adapted versions of the International Prognostic Score (IPI) – the so-called MIPI scores, which incorporate age, ECOG performance status, leucocyte count, lactic dehydrogenase levels and in some variants also the Ki-67 proliferation index – are used in clinical practice [1]. These MIPI scores were built upon data for, and used for prediction of, 5-year survival, with a focus on overall survival (OS). No clinical, laboratory, or biological markers are currently established for prediction of shorter term clinical outcomes.

Pretherapy positron emission tomography/computed tomography after injection of the radiolabelled glucose analogue 2-¹⁸F-fluoro-2-deoxy-d-glucose ([¹⁸F]FDG PET/CT), which enables whole-body in vivo quantification of tumour glucose metabolism, has been shown to provide prognostic information in Hodgkin, diffuse large B cell (DLBCL), follicular and T cell lymphomas in a considerable number of studies, using quantitative parameters including the maximum standardized uptake value (SUVmax), total metabolic tumour volume (TMTV) and total lesion glycolysis (TLG) [3‐9]. In MCL patients, only two studies have investigated the prognostic value of pretherapy SUVmax, with only one of these also including TMTV and TLG [10, 11]. The prognostic value of quantitative measures of [¹⁸F]FDG uptake heterogeneity across the TMTV, as can be provided by advanced radiomic analyses, have not so far been investigated in MCL patients.

Radiomics is an emerging field of research that is concerned with the computer-assisted extraction of quantitative, minable data from diagnostic medical images. Radiomic features include both traditional, first-order features (such as mean and maximum grey-level values), and more sophisticated features such as those that describe different aspects of image texture, which cannot be perceived by the human eye [12]. These image textural features have the potential to allow assessment of tumour heterogeneity [13, 14], which is recognized as a prognostic determinant of survival in different types of cancer [15‐17]. Indeed, several studies in different types of cancer, and using different imaging techniques, have provided data that support the prognostic potential of radiomics [18‐21], especially when processed by artificial intelligence-based machine-learning algorithms.

We therefore aimed to determine (1) whether [¹⁸F]FDG PET-derived radiomic features can predict 2-year progression-free survival (PFS), alone or in combination with clinical, laboratory and biological parameters, using a machine-learning algorithm, and (2) whether the [¹⁸F]FDG PET-based radiomic signature has prognostic value in comparison to, as well as in combination with, the established MIPI scores, in MCL patients receiving CD20 antibody-based immuno(chemo)therapy as first-line systemic treatment.

A total of 107 consecutive patients (35 women and 72 men; mean age 64.5 ± 10.8 years) met the criteria for participation in the study (Fig. 2). [¹⁸F]FDG PET/CT was performed using the Discovery STE scanner in 43 patients, the Discovery 690 scanner in 41 patients, the Discovery 600 and 710 scanners in 10 patients each, and the Discovery ST scanner in 3 patients. Of the 107 patients, 58 were treated with an R-CHOP-based regimen (with additional HiDAC and HDT/ASCT in 42 patients, of whom 10 also received radioimmunotherapy), 38 with R-BENDA or O-BENDA, and 11 with rituximab or ofatumumab monotherapy. After 2 years, progression had occurred in 35 of the107 patients (37.2%), including 8 who had died. The median follow-up period was 52.6 months from registration, the median PFS was 50.4 months, and the median time to progression was 16.3 months. The patients’ baseline clinical, laboratory and biological characteristics are provided in Table 1.

Our study identified two main image-based predictors of 2-year PFS in MCL patients receiving CD20 antibody-based immuno(chemo)therapy: the average glucose metabolism across the entire MTV, as reflected by the SUVmean on pretherapy [¹⁸F]FDG PET/CT scans, and the heterogeneity of glucose metabolism within this MTV, as reflected by Entropy. The combination of the two radiomic features – i.e. the [¹⁸F]FDG PET radiomic signature – may also be useful for outcome prediction: higher SUVmean and higher Entropy appear to be associated with a shorter PFS (Fig. 4).

The three “metabolic risk” categories based on the radiomic signature were superior to the MIPI risk categories in terms of PFS prognostication, regardless of whether or not the Ki-67 proliferation index was considered in the calculation of the MIPI (Table 2, Fig. 4). The performance of the MIPI scores is not necessarily surprising, as OS, and not PFS, was the primary endpoint used in the development of MIPI and MIPI-b [26]. Notably, both MIPI and MIPI-b scores were considerably improved through combination with the [¹⁸F]FDG PET-based metabolic risk. In particular, unlike the MIPI-b, the MIPI-bm not only emerged as a statistically significant predictor of PFS, but was superior to metabolic risk alone (Table 2, Fig. 4), indicating that the best results may be achieved when clinical, laboratory and biological, as well as metabolic information, are integrated in a single model. This is also supported by the results of our machine-learning experiment, in which the addition of ECOG performance status, WBC, LDH level and Ki-67 index to the radiomic features clearly improved 2-year PFS prediction (Fig. 3). Since [¹⁸F]FDG PET/CT is currently recommended for staging and treatment response assessment in patients with MCL by the International Conference on Malignant Lymphoma (ICML) [27] – i.e. it is considered a standard procedure in these patients – information on the metabolic risk represents routine data, which may facilitate its integration into risk assessment in clinical practice.

The prognostic value of pretherapy quantitative [¹⁸F]FDG PET/CT in treatment-naive MCL patients has, to our knowledge, only been investigated in two prior studies. In a series of 81 patients, Karam et al. used the SUVmax to identify groups of MCL patients at risk of shorter survival [10]. The design of their study differed from ours in several ways. First, the SUVmax, which provides information about the single voxel with the highest glucose metabolism within the tumour volume, was the only [¹⁸F]FDG PET-based parameter evaluated, whereas radiomic analysis as performed in our study captures multiple facets of glucose metabolism across the entire metabolic lymphoma volume. Second, patients in the study by Karam et al. were chiefly treated with single-agent or combination chemotherapy (e.g. chlorambucil or CHOP), which does not reflect the present-day therapeutic state-of-the-art for first-line systemic treatment in MCL [29]. Third, three different PET scanners were used, but no correction was performed for the technical differences. Finally, while Karam et al. were able to successfully identify two MCL risk categories using an SUVmax cut-off value of 5, a further subdivision of MCLs with SUV >5 failed; no combination with MIPI scores or individual clinical/laboratory/biological data was attempted. In the second, very recent study, Albano et al. retrospectively evaluated three quantitative [¹⁸F]FDG PET/CT parameters – baseline SUVmax, TMTV and TLG – for outcome prediction in 87 MCL patients, using two different PET/CT scanners, also without applying correction for possible technical differences. These authors found that TMTV and TLG, but not SUVmax, were significantly associated with PFS using two risk categories. Contrary to our own study, neither SUVmean nor radiomic textural features reflecting the heterogeneity of glucose metabolism across the entire tumour, were included in their analysis, and no comparison or combination with MIPI scores was performed [11].

Entropy, a radiomic textural feature derived from the co-occurrence matrix, describes the degree of randomness, or disorder, in the distribution of image voxel grey-level values. Moon et al. recently demonstrated that [¹⁸F]FDG PET/CT-derived Entropy is correlated with the genetic heterogeneity index in lung cancer [13], whereas Choi et al. found that dual-energy CT-derived radiomic features, including Entropy, are also strongly correlated with the pathological heterogeneity index in lung cancer [14]. Entropy extracted from [¹⁸F]FDG PET/CT has recently been used for PFS prediction in patients with high-risk squamous cell carcinoma of the oropharynx after chemoradiation [30], and in patients with lung cancer after EGFR tyrosine kinase inhibitor treatment [31], and has also been found to be associated with failure to respond to third-line systemic treatment in metastatic colorectal cancer [32]. In lymphoma, however, pretherapy [¹⁸F]FDG PET/CT-based Entropy has so far been evaluated for prediction of interim response (i.e. the outcome after two therapy cycles) in paediatric Hodgkin lymphoma [33], and for prediction of disease-free survival and OS in aggressive B cell non-Hodgkin lymphoma (predominantly DLBCL, with a follow-up period of 3–54 months) [34], but did not emerge as a statistically significant marker in either study. Our study is therefore the first to show the value of Entropy on [¹⁸F]FDG PET/CT for outcome prediction in patients with a distinct histological lymphoma subtype, a finding that could be due to the fact that MCL is a “genomically unstable” tumour that may be associated with marked (sub)clonal heterogeneity, and also with heterogeneity between different topographic sites in the same patient, and with modulation of the initial mutational profile during disease progression [35].

We used PFS instead of OS as the clinical endpoint, which is contrary to many lymphoma studies investigating prognostic features. This strategy was chosen because MCL is a less common lymphoma subtype, and hence smaller patient populations and shorter follow-up periods may have to be used to obtain a sufficiently large number of cases with the event of interest. While OS is the established outcome, PFS has been recognized by the Food and Drug Administration as a valid surrogate endpoint in MCL and other haematological malignancies [36]. In a recent analysis of multiple randomized trials in DLBCL, including a total of 7,507 patients, PFS was significantly correlated with OS, supporting its use as a surrogate marker (https://​www.​fda.​gov/​Drugs/​DevelopmentAppro​valProcess/​DevelopmentResou​rces/​ucm613636.​htm).

Our study had some limitations. The most obvious limitations were the retrospective design and the modest cohort size. However, this was a hypothesis-generating study, as [¹⁸F]FDG PET-derived radiomic features have not been previously evaluated for outcome prediction and prognostication in MCL patients. Furthermore, with a sample size of 107 patients, this is the largest study on this topic at present. Our approach involving the combination of radiomic features and MIPI scores – i.e. using the three categories of metabolic risk to modify MIPI scores – was exploratory. However, MIPI scores were clearly improved by applying this strategy, and in addition, the combination of the (continuous) radiomic feature values and the clinical, laboratory and biological parameters (which were also used for calculation of the MIPI scores), using a machine-learning algorithm, showed the same trend: best results were achieved when all parameters were integrated into a single model. We used an MLP neural network for outcome prediction, which is a universal function approximator with the ability to model any type of regression or classification problem [37]. While MLP networks are well established in the machine-learning community as powerful prediction algorithms [38], it is possible that even more advanced, deep machine-learning techniques such as convolutional neural networks (CNN), with their larger numbers of hidden layers and their interconnection between neurons within the same layer, may have performed even better. However, CNNs are mainly intended for use with large datasets (“big data”). In a patient population such as our own, however, their complexity would have increased the probability of “overfitting”, i.e. loss of generalizability of the model, in our case for PFS prediction [39].

In conclusion, an [¹⁸F]FDG PET radiomic signature comprising SUVmean and Entropy has prognostic value in MCL and may be useful for predicting early tumour progression. This metabolic risk reflected by radiomic features can be integrated into MIPI scores and may possibly improve risk stratification in MCL. Further studies are warranted to validate these findings in external cohorts.