Functional Imaging using Radiomic Features in Assessment of Lymphoma

Abstract

Lymphomas are typically large, well-defined, and relatively homogeneous tumors, and therefore represent ideal targets for the use of radiomics. Of the available functional imaging tests, [18F]FDG-PET for body lymphoma and diffusion-weighted MRI (DWI) for central nervous system (CNS) lymphoma are of particular interest. The current literature suggests that two main applications for radiomics in lymphoma show promise: differentiation of lymphomas from other tumors, and lymphoma treatment response and outcome prognostication. In particular, encouraging results reported in the limited number of presently available studies that utilize functional imaging suggest that (1) MRI-based radiomics enables differentiation of CNS lymphoma from glioblastoma, and (2) baseline [18F]FDG-PET radiomics could be useful for survival prognostication, adding to or even replacing commonly used metrics such as standardized uptake values and metabolic tumor volume. However, due to differences in biological and clinical characteristics of different lymphoma subtypes and an increasing number of treatment options, more data are required to support these findings. Furthermore, a consensus on several critical steps in the radiomics workflow –most importantly, image reconstruction and post processing, lesion segmentation, and choice of classification algorithm– is desirable to ensure comparability of results between research institutions.

1. INTRODUCTION – LYMPHOMA AND IMAGING

Lymphomas are a heterogeneous group of neoplasms that belong to the larger family of hemato-oncological malignancies. The traditional division of lymphomas into Hodgkin and Non-Hodgkin lymphomas is still in use today, as Hodgkin lymphoma is unique in terms of histology, treatment options and clinical prognosis. Non-Hodgkin lymphomas (NHL) show a high degree of clinical, pathologic and prognostic heterogeneity, partly overlap with specific types of leukemia, and can be roughly divided into aggressive and indolent subtypes depending on their rates of proliferation and clinical course [1].

Diffuse large B-cell lymphoma (DLBCL) is the most common type of NHL, and also the most common aggressive NHL, whereas follicular lymphoma is the second most common NHL, and the most common indolent NHL – both belong to the group of B cell NHLs that is more common than the T cell and natural killer cell group. Other more frequently seen subtypes are marginal zone lymphoma (MZL), mantle cell lymphoma (MCL), small cell lymphocytic lymphoma (SLL), angioimmunoblastic T cell lymphoma (AITL) and peripheral T cell lymphoma (PTCL) [2].

Imaging for characterization, treatment response assessment and outcome prognostication is challenging due to the above described pathohistological and clinical heterogeneity within the lymphoma family, and an increasing array of treatment options that comprise traditional chemotherapy regimens, radiation therapy, immune therapies (including established and novel monoclonal antibodies, such as rituximab, and PD-1 checkpoint inhibitors; immunomodulatory drugs such as lenalidomide; and CAR T-cells), small molecule drugs such as the tyrosine kinase inhibitor ibrutinib, and stem cell transplantation.

Despite the above described complexity, [18F]FDG-PET (2-[18F]-fluoro-2-deoxy-D-glucose), which captures glucose metabolism, has established itself as the technique of choice for the majority of lymphomas, not just for the detection, but also for response assessment [3, 4]. There is also increasing evidence that [18F]FDG-PET may have a role for treatment stratification [5–7] and outcome prognostication, the latter in particular with (semi-) quantitative metrics such as standardized uptake values (SUV), total metabolic tumor volumes (MTV) and total lesion glycolysis (TLG) [8–14]. For example, a large, recently published study by Vercellino et al. in 310 DLBCL patients suggested that, in this lymphoma subtype, baseline total MTV represents and independent prognostic factor for both overall survival (OS) and progression-free survival (PFS), regardless of the actual response to treatment. An exception to this widespread application of [18F]FDG-PET are some indolent lymphomas, such as the MZL and SLL that frequently show low, or even no FDG uptake or contrast relative to background uptake; and central nervous system (CNS) lymphomas that may be difficult to capture due to the physiologic cerebral glucose metabolism, in particular when small. For such indolent and CNS lymphomas, morphologic imaging is currently recommended according to the Lugano guidelines [3]. The only functional –although not metabolic– imaging technique that has been more extensively evaluated for the use in lymphomas is diffusion-weighted MRI, which enables indirect assessment of cell density based on tissue diffusivity, and captures treatment induced cell necrosis within the first 72 hours of treatment initiation [15, 16]. DWI is based on the restriction of water movement in extracellular space that is compressed due to the increased cell number and/or size within tumor tissue; apparent diffusion coefficients (ADCs) can be extracted from DWI as a quantitative parameter.

2. RADIOMICS FOR LYMPHOMA: RATIONALE AND TECHNIQUE

2.1. Clinical rationale

As a “systemic” malignancy that, in the majority of cases, lacks a “primary tumor”, lymphoma manifestations are found at different anatomic sites that are not limited to lymph node stations across the body, but also include extranodal sites, i.e., solid organs, soft tissues and bone marrow. Biopsies are usually only performed for a single lymph node and/or a single extranodal site, as well as (with the exception of Hodgkin lymphoma) the bone marrow at the level of the iliac crest [3]. This means that biopsies or lesion resection may not adequately reflect heterogeneity across the entire tumor volume, capturing only information from a very small number of sites, or even a single site. Tumor heterogeneity has, however, been shown to be a relevant prognostic factor in different cancers [17–18]. Currently used visual and quantitative [18F]FDG-PET metrics, such as maximum and mean standardized uptake values (SUVs), however, fail to capture heterogeneity, as they only provide very basic histogram features.

Here, more sophisticated radiomic features that capture complex mathematical patterns in the spatial distribution of signal intensity values of voxels have demonstrated their potential to assess tumor heterogeneity non-invasively [19–22]. This information could potentially be useful for prediction of treatment response, as well as clinical outcome prognostication. Another possible application could be the early detection of transformation from indolent to more aggressive lymphoma subtype at PET/CT follow-up. For CNS lymphoma, lesion characterization, and in particular, image-based discrimination between lymphoma and primary malignancies of the brain would have clinical impact – after all, CNS biopsies are technically more challenging and pose a greater risk of complications than, e.g., superficial lymph node biopsies.

2.2. Technique and challenges

From a technical point of view, radiomics for lymphoma does not differ from radiomics for other cancers –the same principles of feature extraction based on, e.g., the grey-level co-occurrence matrix (GLCM), run-length matrix (GLRLM), size-zone matrix (GLSZM), neighborhood gray-tone difference matrix (NGTDM), or neighborhood gray-level dependence matrix (NGLDM) [23]. A comprehensive list of radiomic features including mathematical equations can be found in the white paper of the IBSI (Image Biomarker Standardisation Initiative) [24].

Other general principles for performing meaningful radiomics research, such the use of pre-defined, homogeneous imaging protocols, or, if the latter cannot be achieved (as is frequently the case in multicentric retrospective studies), image resampling and signal intensity discretization, are also applicable, or the use of phantoms to be able to standardize radiomic feature values [25]. This is necessary because it has been demonstrated that variations of image acquisition parameters, such as repetition and echo times for MRI, and number of iterations and reconstruction algorithm for PET, and even more importantly, spatial resolution (voxel size), have a considerable impact on (second- and higher-order) radiomic features [26–29]. For heterogeneous datasets, post-reconstruction harmonization, for instance using the ComBat technique, also represents a possible means to correct for systematic differences in features values caused by the use of different acquisition parameters [30]. Clearly, the effectiveness of such post-processing techniques depends on the degree of variation between protocols. No meaningful correction is presently available for features extracted from images that are degraded by motion artifacts, even if they are minor – while the human eye can partly compensate for such minor artifacts, they can have major effects on image texture. Consequently, high image quality is another prerequisite for radiomics studies.

With regard to lesion segmentation most lymphomas, especially in the body, may be regarded as more straightforward than many other cancers: the characteristic lymph nodes are clearly enlarged, well-defined as single lesions or lymph node bulks/masses, respect organ boundaries (as opposed to diffuse infiltration seen, for instance, in pancreatic ductal adenocarcinoma), and necrosis is less common.

3. LYMPHOMA DIAGNOSIS

3.1. CNS lymphoma

Given the above described limitations of [18F]FDG-PET for CNS imaging, only a single PET radiomics study, but three studies using DWI radiomics were performed so far (Table 1). All of these studies focused on the differentiation between primary CNS lymphoma (PCNSL) and glioblastoma multiforme (GBM), a topic that is particularly challenging when PCNSL shows atypical features such as intralesional haemorrhage, necrosis, or heterogeneous contrast-enhancing components.

TABLE 1.

Radiomics studies focusing on lymphoma diagnosis / detection

	Purpose	Imaging	Patients	Training/Test	Segmentation	Features
Kong et al	PCNSL vs. GBM	[18F]FDG-PET	77	No	Manual	Histogram, GLCM, GLRLM, GLSZM, NGTDM
Kim et al	PCNSL vs. GBM	ADC maps (+other MR sequences)	143	Yes (internal)	Semi-automatic	Shape, Histogram, GLCM, GLSZM
Kang et al.	PCNSL vs. GBM	ADC maps	196	Yes (internal + external)	Semi-automatic	Shape, Histogram, GLCM, GLRM, Wavelet
Yun et al.	PCNSL vs. GBM	ADC maps (+ CE-T1 MRI)	195	Yes (internal + external)	Semi-automatic + Automatic	Histogram, GLCM, GLRLM, Wavelet, CNN-based
Ou et al.	Lymphoma vs. breast cancer	[18F]FDG-PET and CT	44	Yes	Manual	Shape, Histogram, GLCM, GLRLM, GLSZM, NGTDM
Zhu et al.	Lymphoma vs. renal cell cancer	[18F]FDG-PET	38	No	Manual	Shape, Histogram, GLCM, GLRLM, GLSZM, NGTDM
Aide et al.	DLBCL: Bone marrow involvement	[18F]FDG-PET	82	No	Semi-automatic	Histogram, GLCM, GLSZM
Mayerhoefer et al.	MCL: Bonw marrow involvement	[18F]FDG-PET	97	Yes (internal)	Semi-automatic	Histogram, GLCM

NHL, Non-Hodgkin lymphoma; OS, overall survival; DFS, disease-free survival; PFS, progression-free survival; GLCM, gray-level co-occurrence matrix; GLRLM, gray-level run-length matrix; GLSZM, gray-level size-zone matrix; NGTDM, neighborhood gray-tone difference matrix; NGLDM, neighborhood gray-level dependence matrix

Kong et al. retrospectively included 77 patients (24 with PCNSL and 53 with GBM) who had undergone preoperative [18F]FDG-PET and subsequently underwent surgical resection or biopsy [31]. Imaging was performed on a single PET/CT scanner, and tumors were segmented manually. Radiomic analysis was performed on three sets of images: unmodified PET; PET with normalization of tumor voxels using contralateral cortical FDG uptake; and PET with normalization of tumor voxels using contralateral mean brain FDG uptake. SUV, MTV, TLG, and 193 histogram, GLCM, GLRLM, GLSZM and NGTDM features were calculated in 3D. Mean AUCs obtained using 5-fold cross-validation were calculated for features that differed significantly between PCNSL and GBM, were stable, and showed better discriminatory performance than SUVmax. 13 individual radiomic features met these criteria and provided AUCs 0.971–0.998. No feature combinations were evaluated.

Kim et al. investigated the value of radiomic features extracted from different MRI sequences, including DWI and contrast-enhanced sequences obtained with a single 3T MRI scanner model available at two institutions; acquisition protocols for the relevant MR sequences were also identical [32]. A total of 143 patients were retrospectively included and divided into a training (“discovery”) set (49 with GBM and 37 with PCNSL) and validation set (29 with GBM and 28 with PCNSL). Out of 127 radiomic features (histogram, GLCM, GLSZM, and shape-based) calculated after semi-automatic segmentation using a level tracing algorithm, two types of feature selection/reduction technique were applied: a minimum redundancy maximum relevance (mRMR) algorithm, and a least absolute shrinkage and selection operator (LASSO) algorithm with ten-fold cross-validation, based on the combined data of all sequences, respectively. A logistic regression, a support vector machine (SVM), and a random forest (RF) algorithm were used for classification. Using a subset of 15 selected features (including 4 ADC map-based histogram and GLCM features), AUCs of >0.95 were achieved in both the training and validation set for differentiation between of PCNSL and GBM.

Kang et al. used a three-step approach for DWI radiomics-based differentiation of PCNSL and GBM [33]. In their retrospective study, they used 112 patients (42 PCNSL and 70 GMB) as the training dataset, 42 patients (21 PCNSL and 21 GBM) as internal validation dataset, and another 42 patients (14 PCNSL and 28 GBM) obtained at another center for external testing that used the same 3T MR scanner model, but slightly different sequence parameters. Semi-automatic lesion segmentation was performed using a region-growing algorithm, and following ADC map spatial resampling and signal intensity normalization, 1618 features (shape, histogram, GLCM, GLRM, and, as the largest group 1432 wavelet features) were calculated. Notably, 12 different techniques for feature selection were used, including mRMR, mutual information and Fisher scores; and eight classification techniques, including a k-nearest neighbor and a random forest algorithm, linear discriminant analysis, and SVM. Excellent results were achieved using a combination of mRMR and the random forest classifier, yielding comparable results in the training, validation and test datasets (AUCs 0.98 and 0.94, respectively), thus performing better than human evaluation (AUC, 0.83–0.91) and the routine MRI metrics cerebral blood volume (AUC 0.91) and ADC (0.79).

Finally, Yun et al. also used a training dataset (50 PCNSL and 73 GBM), and internal validation set (12 PCNSL and 18 GBM) and an external test dataset (14 PCNSL and 28 GBM) for establishing a robust network for MRI radiomics-based discrimination between the two tumor types [34]. Following signal intensity normalization and resampling, ADC maps and contrast-enhanced MR images were analyzed, and based on semi-automatic segmentation algorithm (threshold coupled with a region-growing algorithm), 936 radiomic features were extracted (histogram, GLCM, and GLRLM, before and after wavelet decomposition). A variety of supervised machine learning techniques, including SVM and random forest algorithms, were used for classification, in combination with three feature selection techniques. Furthermore, a multilayer perceptron neural network (MLP-NN) with ten-fold cross-validation without previous feature selection was applied. Finally, in addition to these classic radiomics approaches, automatic image analysis, feature extraction and classification were also performed by a CNN. MLP-NN provided the best performance overall, with results that were comparable in the internal validation and external test datasets, and with an AUC of 0.95 in the external test dataset, were comparable to those achieved by human expert readers (AUCs up to 0.93); notably, CNN performed poorly (AUC, 0.49).

3.2. Body lymphoma

Despite the fact that lymphoma involvement is far more common in the body, and in particular the lymph nodes, than involvement of the CNS, comparatively few radiomics studies for body lymphoma exist at present. Compared to CNS lymphoma, cohorts are also smaller, and the technical approach in terms of features classes, segmentation techniques, and classifiers is also less complex.

3.2.1. Solid organs

Extranodal lymphoma manifestations in different organs can mimic other, usually primary cancers. In current clinical practice, the diagnosis is established by biopsies, which, depending on the anatomic site, carry a small but non-negligible risk of bleeding or other organ damage. Therefore, radiomics may possibly become an alternative to differentiate between lymphoma and other primary cancers, especially when the characteristic marked lymphadenopathy is not present. Compared to PCNSL, which histologically corresponds to DLBCL, histological variety is far greater in the body, making the comparison between results of studies in different lymphoma subtypes difficult (Table 1).

Ou et al. investigated whether [18F]FDG-PET/CT radiomics can distinguish between primary breast cancer and breast tissue lymphoma, which accounts for approximately 0.5% of breast malignancies [35]. Forty-four previously untreated patients –25 with breast cancer and 19 with lymphoma– were retrospectively included. 3D tumor segmentation was performed manually, and following spatial resampling and intensity rescaling and discretization, standard PET and CT metrics (SUVs, MTV, TLG, and CT density values) as well as histogram and texture features were extracted independently from PET and CT images –one of the currently rare examples of PET/CT radiomics studies where features were extracted from both the PET and the CT component. Despite its small size, the cohort was divided into a training dataset (80%) and a validation dataset (20%); LASSO was used for feature selection. LDA results showed that both for PET and CT, combination of respective radiomic features with clinical parameters and standard metrics provided the best discrimination, with AUCs of 0.81 and 0.76, respectively. Notably, however, combination of PET and CT radiomic features was not attempted.

In a cohort of 38 patients Zhu et al. evaluated the ability of [18F]FDG-PET/CT radiomics to differentiate between renal lymphoma manifestations (20 patients) and renal cell carcinoma (18 patients) [36]. Both PET and CT images were analyzed, and after manual lesion segmentation, 45 radiomic features (including both conventional, texture and shape features) were analyzed, first independently, and then on a per-feature class basis using binary logistic regression analysis. Twelve individual radiomic texture features and two histogram features showed AUCs of >0.7. Feature combinations of the different classes revealed clearly higher AUC, of up to 1.0 for GLRLM. However, due to the limited size of the two cohorts, no division of the patient population into a training and a validation dataset was possible, which clearly increases the risk of overfitting; therefore, more data are needed to further explore this potential application of [18F]FDG-PET/CT radiomics.”

3.2.2. Bone marrow

Image-based assessment of bone marrow involvement is a topic of interest to clinicians, because in case of a multi-focal or diffuse spread, it fulfills the criterion for Lugano stage IV [3]. According to the Lugano guidelines, bone marrow remains the only organ for which –with the exception of Hodgkin lymphoma– no imaging test, but biopsy remains the standard of care to rule out lymphoma [3].

Aide et al. retrospectively investigated baseline [18F]FDG-PET radiomics-based prediction of bone marrow involvement in 82 patients with DLBCL [37]. 3D radiomic features (histogram, GLCM, GLSZM) were extracted after CT (Hounsfield unit) threshold-based segmentation of the skeleton coupled with manual exclusion of physiologic FDG uptake. No separation into a training and validation set was performed, and radiomic features were tested individually. For assessment of bone marrow involvement, histogram features were superior to texture features, with AUCs of up to 0.82 (for the feature Skewness) versus AUCs up to 0.75; notably, however, a biopsy reference standard was not available in all patients, and there was no separate validation dataset.

In a retrospective analysis of 97 patients with MCL that were split into training (70%) and validation (30%) datasets, we recently investigated whether baseline [18F]FDG-PET GLCM features can improve SUVs for assessment of bone marrow involvement [38]. Notably, MCL can either have an indolent course, in which case bone marrow FDG uptake is low; or, more frequently, an aggressive course, where it shows substantial FDG uptake [39]. Only the bony pelvis was segmented semi-automatically (41% SUVmax threshold and manual adjustment) so as not to include areas of degenerative uptake in the spine, and because biopsies that served as the reference standard in all cases were also obtained at this site. MLP-NN showed that principal radiomic components generated from SUVs and the 16 GLCM features provided better assessment of bone marrow involvement (AUCs of up to 0.73) than SUVs alone (AUCs of up to 0.66); even better results were observed when the radiomic signature was combined with white blood count (WBC) and lactate dehydrogenase (LDH) (AUCs of up to 0.81). If the threshold of marrow involvement was raised to 5% or 10%, results for the radiomics signature (AUCs up to 0.84 and 0.85) and their combination with laboratory values (up to 0.84 and 0.87) increased.

4. LYMPHOMA OUTCOME PREDICTION AND PROGNOSTICATION

With the large spectrum of possible treatments for lymphoma, outcome prognostication and prediction of disease status are of considerable interest for clinicians. Traditional scores used in lymphomas, such as the International Prognostic Index (IPI, used in DLBCL) and its modification for follicular lymphoma (FLIPI) and MCL (MIPI) comprise laboratory findings (e.g., white blood count (WBC) and LDH), clinical performance status, and depending on the variant, also biological data such as the ki-67 proliferation index [40]. Addition of [18F]FDG-PET metrics to these clinical scores, or even replacement of the latter, has been discussed in an increasing number of studies. So far, however, PET metrics we limited to SUVs, MTV and TLG, whereas more advanced radiomic features were only studied in a small number of publications, which are summarized below (Table2).

TABLE 2.

[18F]FDG-PET radiomics studies focusing on clinical outcome prognostication

	Purpose	Imaging	Patients	Training/Test	Segmentation	Features
Parvez et al.	Aggressive NHL: treatment response, OS, DFS	[18F]FDG-PET	66	No	Semi-automatic	Shape, Histogram, GLCM, GLRLM, GLSZM, NGTDM
Milgrom et al.	Hodgkin: Refractory disease	[18F]FDG-PET	251	Yes (internal)	Semi-automatic	Shape, Histogram, GLCM
Lue et al.	Hodgkin: OS, PFS	[18F]FDG-PET	42	No	Semi-automatic	Histogram, GLCM, GLRLM, GLSZM, NGTDM, NGLDM, Wavelet
Mayerhoefer et al.	MCL: PFS	[18F]FDG-PET	107	Yes (internal)	Semi-automatic	Histogram, GLCM
Aide et al.	DLBCL: PFS, OS	[18F]FDG-PET	82	No	Semi-automatic	Histogram, GLCM, GLSZM
Aide et al.	DLBCL: EFS	[18F]FDG-PET	132	Yes (internal)	Semi-automatic	Histogram, GLCM, GLSZM
Wand et al	Natural killer-/T-cell lymphoma: PFS, OS	[18F]FDG-PET	110	Yes (internal)	Semi-automatic	Shape, Histogram, GLCM, GLRLM, GLSZM, NGLDM

NHL, Non-Hodgkin lymphoma; OS, overall survival; DFS, disease-free survival; PFS, progression-free survival; GLCM, gray-level co-occurrence matrix; GLRLM, gray-level run-length matrix; GLSZM, gray-level size-zone matrix; NGTDM, neighborhood gray-tone difference matrix; NGLDM, neighborhood gray-level dependence matrix

4.1. Hodgkin lymphoma

Milgrom et al. retrospectively analyzed [18F]FDG-PET data of 251 patients with classical HL stage I-II to predict refractory or relapsed disease [41]. Images were obtained with four scanners from the same vendor, and moderate dose variations; no post-reconstruction correction was applied. The analysis comprised both standard and more sophisticated histogram features, GLCM features and basic shape features, extracted from the mediastinal tumor volume on the one hand, and the total tumor volume on the other hand, segmented by a combination of manual and threshold-based delineation. The cohort was split into training and validation datasets, and SVM and unsupervised hierarchical clustering were used to predict residual/refractory disease within 90 days after end of first-line treatment. Based on the five most predictive features of the mediastinal tumor volume, AUCs of 0.95 for the radiomic approach versus 0.78 for MTV and TLG, and just 0.65 for SUVmax were achieved; whereas total tumor volume PET features were not predictive of outcome.

Lue et al. retrospectively analyzed single-scanner pre-therapeutic [18F]FDG-PET scans of 42 HL patients that subsequently underwent chemo- or radiochemotherapy [42]. This study also focused on survival prognostication –in this case, overall survival (OS) and progression-free survival (PFS)– by means of overall 450 3D radiomic features extracted from original/normalized and wavelet-decomposed images. No separate validation dataset was used. In their multivariate Cox regression analysis, only a single GLRLM feature (intensity non-uniformity) remained prognostic for both OS and PFS, and a single histogram features (SUV kurtosis) was prognostic for PFS alone. Of note, MTV, a metric that has shown prognostic value in several studies in HL patients, was neither prognostic for OS or PFS.

4.2. Non-Hodgkin lymphoma

Parvez et al. evaluated [18F]FDG-PET standard and advanced radiomic features (histogram, shape, GLCM, GLRLM, GLSZM and NGTDM) for treatment response prediction and prognostication of OS and disease-free survival (DFS) in 66 patients with aggressive NHL [43]. The vast majority of patients belonged to the DLBCL subtype and was therefore subsequently treated with R-CHOP. Patients underwent PET/CT on a single scanner, eliminating the need for correction. Notably, only selected tumor sites with high FDG uptake were included in the 3D analysis. Individual features, but no features combinations were investigated. While several radiomic texture features were prognostic for DFS, only kurtosis was prognostic for OS, in addition to SUV and TLG variants. Contrary to MTV, no texture feature proved useful for first-line therapy outcome prediction.

We retrospectively evaluated baseline [18F]FDG-PET radiomic features of patients with MCL, which shows a variable course under standard treatment, for treatment response assessment [44]. The cohort consisted of 107 patients, which were scanned with five different PET/CT scanners from the same vendor (General Electric Discovery series), but otherwise identical image acquisition parameters; nevertheless, we performed ComBat post-reconstruction harmonization, as described by Orlhac et al [29], to account for the differences between scanners. Apart from SUVs, MTV, and TLG, only 16 MTV-based GLCM features were evaluated for PFS prognostication, in an attempt to keep the data dimensionality low to avoid overfitting; the cohort was split into a training and an internal validation cohort (70:30%). GLCM Entropy and SUVmean emerged as the two radiomic features predictive for 2-year PFS. The combination of Entropy and SUVmean (AUC of 0.73 for 2-year PFS), termed “metabolic risk”, was prognostic for PFS, and their integration with clinical MIPI scores, and especially MIPI-b, clearly improved the 3-category risk model.

Aide et al. performed two [18F]FDG-PET radiomics studies in DLBCL patients, both focusing at least in part on survival prognostication. In their above described 2018 study (see section 3.2.1) that utilized radiomic features extracted from the bone marrow, these authors also evaluated the prognostic value of Skewness (i.e. the feature that was most predictive of bone marrow involvement) for OS and PFS prognostication [37]. While Skewness turned out to be the only prognostic factor for PFS in the multivariate analysis, only the IPI score, but no radiomic feature, was an independent predictor of OS.

On the other hand, in their 2020 study, Aide et al. analyzed the value of pre-therapeutic [18F]FDG-PET radiomic features, for 2-year EFS prediction (which was achieved by 77% of the 132 DLBCL patients) and EFS prognostication [45]. The same features classes as in their bone marrow study were extracted, in this case exclusively from each patient’s largest lesion, rather than from the entire MTV. Also, the cohort was split into a training dataset (105 patients) and a validation dataset (27 patients).While several GLCM features (AUCs up to 0.64) and GLSZM features (AUCs up to 0.69) were predictive of 2-year EFS, the GLSZM feature Long-Zone High-Grey Level Emphasis was the only parameter prognostic for 2-year EFS in the multivariate analysis.

Very recently, Wang et al. retrospectively compared [18F]FDG-PET metabolism-based models (SUVmax, MTV, and TLG, as well as clinical parameters such as ECOG score and international prognostic index) to radiomics-based models (radiomic signature combined with clinical parameters) in terms of PFS and OS prediction in 110 patients with nasal-type extranodal natural killer/T-cell lymphoma scheduled to receive chemo- or radiochemotherapy [46]. PET/CT was performed on a single scanner, but nevertheless, spatial and intensity resampling as well as intensity discretization were performed. The cohort was split into 82 and 28 patients in the training and validation cohorts, respectively. Correlation coefficients and LASSO coefficients were used for feature selection to generate a radiomic signature based on 38 shape, histogram, second and higher order features. Notably, results showed that the metabolic model performed better than the radiomics-based model for prognostication of OS, with comparable results in the training dataset (C-indices of 0.83 vs. 0.82), but poorer results for the radiomics-based model in the validation dataset (C-indices of 0.75 and 0.63). For PFS prognostication, the metabolism-based model was inferior to the radiomics-based model in the training cohort (C-indices of 0.75 and 0.81) but superior in the validation cohort (C-indices of 0.69 and 0.59). These results clearly demonstrate that radiomics is not necessary superior to traditional (semi-)quantitative imaging features.

5. SUMMARY

The literature on functional imaging radiomics for lymphoma is still scarce. Nevertheless, the published research in this field highlights the two main areas of clinical interest: distinguishing lymphoma from other cancers, especially in the CNS, where PCNSL may exhibit atypical features that resemble those of GBM; and lymphoma outcome prognostication. For CNS lymphoma, MRI / DWI radiomics studies currently outnumber PET studies, but this may change as more specific tracers such as [18F]fluoroethyltyrosine (FET) or [68Ga]Pentixafor, which unlike [18F]FDG do not show a relevant degree of physiological background uptake in the CNS, gain more acceptance. Applications of radiomics in the CNS may also possibly extend towards treatment response assessment, since the differentiation between post-radiation necrosis and residual viable tumor, as well as between post-treatment pseudo-progression and true progression, represent common diagnostic dilemmata.

With regard to survival prognostication, pre-therapeutic [18F]FDG-PET radiomics may provide complementary information to currently used PET metrics such as SUVs and MTV, especially in lymphomas with a high degree of clonal instability, such as mantle cell lymphoma, where aggressive and indolent manifestations may be found in the same patient. In such patients a whole-tumor measure of metabolic heterogeneity may indeed provide complementary information. Evaluation of anatomic sites with diffuse infiltration by lymphoma cells, such as the bone marrow, where MTV is not applicable and SUVs have yielded poor results especially in indolent NHL, may represent another possible application. However, contrary to SUVs and MTV that are already well- established or on the verge being introduced into clinical practice (in particular for outcome prognostication, and based on it, treatment stratification), more data is clearly needed on the respective roles of radiomics in different aggressive and indolent lymphoma subtypes, and under different treatment regimens, ideally in the form of prospective studies which have so far not been undertaken. Here, conformance with widely accepted recommendations for radiomics research – including trial registration, imaging protocol homogeneity, internal and external validation datasets, and use of feature dimensionality reduction to reduce the risk of overfitting – is essential. Provided that these quality criteria are met, there is also a need to report and publish negative results, so as to provide a realistic picture of the added value of radiomics in biomedicine, or lack thereof, as recently pointed out by Buvat and Orlhac [47].

HIGHLIGHTS.

• Functional MR radiomics may distinguish between CNS lymphoma and glioblastoma.

• Baseline [18F]FDG-PET radiomics may enable lymphoma outcome prognostication.

• There is a lack of prospective trials of functional imaging radiomics for lymphoma.