Introduction

Diffuse large B-cell lymphoma (DLBCL) is the most common type of lymphoma [1]. As a remarkably heterogeneous disease, DLBCL is known for its considerable variation in clinical behaviour, response to therapy, and prognosis [2]. Chemoimmunotherapy with R-CHOP (rituximab, cyclophosphamide, doxorubicin, vincristine, and prednisone) therapy is now used as the standard therapy for DLBCL [3]. However, 30-50% of patients ultimately relapse and die as a result of the disease [4, 5]. Therefore, an accurate identification of patients who will not respond well to standard regimens at diagnosis is needed to guide new therapeutic strategies. The International Prognostic Index (IPI) and its successor (NCCN-IPI) are prognostic scoring systems that predict survival. However, it is difficult to accurately identify DLBCL patients who are likely to relapse, especially those with a poor long-term prognosis, to individualize therapy [6,7,8].

Volume metabolic parameters derived from [18F]-FDG PET/CT, such as metabolic tumour volume(MTV) and total lesion glycolysis (TLG), are powerful indices for assessing tumour burden and risk stratification in DLBCL patients [9,10,11,12,13]. However, these metabolic parameters are limited in their ability to reflect tumour heterogeneity, which ultimately limits their contribution to predicting cancer progression, therapeutic resistance, and survival outcomes [14, 15]. Machine learning advancements have enabled clinicians to use radiomics to extract implicit features from images to assess intratumoural biological heterogeneity and assist in clinical outcome prediction [16, 17]. Several studies [18,19,20,21,22,23] have explored the prognostic value of PET/CT radiomics in lymphoma, including Hodgkin lymphoma (HL), nasal-type extranodal natural killer/T cell lymphoma (ENKTL), mantle cell lymphoma (MCL), and DLBCL. In general, these studies used one machine learning model to select and develop radiomic signatures, e.g. the multilayer perceptron neural network combined with logistic regression analysis [19], the least absolute shrinkage and selection operator regression (LASSO) model [20, 22], and the logistic regression model [23]. However, there is no model that works well for all kinds of data. To address this limitation, cross-combinations of classification and selection methods based on radiomics features have recently appeared and have proven effective in improving classification and diagnostic performance [24, 25]. However, it is still unknown whether these methods can be applied to radiomic images for DLBCL. In this paper, we explored the PET radiomics features using a cross-combination of machine learning models to assess risk in DLBCL patients. Furthermore, we established and externally validated models for predicting survival outcome in routine clinical practice, allowing for its clinical translation.

Materials and methods

Patient data collection

A total of 273 patients from Nanjing Drum Tower Hospital, the Affiliated Hospital of Nanjing University Medical School, were included as the training cohort. To examine the general applicability of nomograms, an independent external validation cohort of 110 patients from Jiangsu Province Hospital, the First Affiliated Hospital of Nanjing Medical University, was included. Ethical approval was obtained for this retrospective analysis, and the requirement for written consent from patients was waived. The workflow of patient selection is shown in Fig. 1. The inclusion criteria were as follows: (I) pathologically confirmed DLBCL, (II) pretreatment [18F]-FDG PET/CT scan was performed, and (III) treatment with R-CHOP-like regimens with curative intent was carried out. Patients with incomplete medical data, those lost to follow-up, and those with a previous history of cancer were excluded. Clinical data (sex, age, B symptoms, Eastern Cooperative Oncology Group performance status (ECOG PS), NCCN-IPI index, lactate dehydrogenase (LDH) level, Ann Arbor stage, and immunohistochemistry result) were obtained from medical records. Participants were divided into two groups: those with a low/low-intermediate risk (NCCN-IPI of <4) and those with a high-intermediate/high risk (NCCN-IPI of ≥4) [8]. The follow-up data were obtained through electronic medical records and telephone interviews. Progression-free survival (PFS) and overall survival (OS) were chosen as endpoints for evaluating the prognosis of DLBCL patients. PFS was calculated as the time between diagnosis and the first relapse, progression, death from any cause, or last follow-up. OS was calculated as the time between diagnosis and death from any cause or last follow-up.

Fig. 1
figure 1

Flow chart of participant selection

Full size image

PET/CT Scanning Protocol

All patients underwent PET/CT scans with one of the following systems: Gemini GXL (Philips Corp, Netherlands) for the training cohort and Biograph 16 PET/CT scanner (Siemens Health care, Erlangen, Germany) for the external validation cohort, based on standard clinical scanning protocols. All patients fasted for at least 6 hours before scans, resulting in blood glucose levels under 8.7 mmol/L. After 6 hours of fasting (no oral or intravenous fluids containing sugar or dextrose), 185–370 MBq of [18F]-FDG (5.18 MBq/kg) was administered intravenously. Each patient was weighed to determine the standardized uptake value (SUV) prior to each scan. Whole-body PET/CT scans (from the base of the skull to the upper thigh) were performed 60 minutes after the radiopharmaceutical injection. Emission data were acquired for 2 minutes in each bed position. CT acquisition data were used for attenuation correction, and corrected PET images were reconstructed using ordered-subset expectation maximization (OSEM). Images were reconstructed with standard 4 × 4 × 4 mm3 voxels using 3 iterations of 33 subsets.

VOI drawing and radiomics processing

PET images were uploaded to 3D Slicer software (version 4.8.0; http://www.slicer.org). The Grow-Cut algorithm in 3D Slicer was used to semiautomatically generate the volume of interest (VOI). Two physicians manually adjusted the VOI to ensure that the measurement was reliable. If there was a discrepancy, the VOIs were reviewed and determined by a senior nuclear medical scientist. The 3 largest lesions were chosen as the targets for radiomic feature extraction. For the three largest VOIs, we summed the radiomic size and shape features, while all other features were averaged for each VOI [26]. Feature extraction was performed based on PET images by the built-in open-source PyRadiomics package (http://www.radiomics.io/pyradiomics.html) in 3D Slicer. Before feature extraction, we preprocessed the PET images, including resampling all images to a voxel size of 1×1×1 mm3 using bilinear interpolation and discretization with a fixed bin count of 128, as recommended by [27]. To obtain the comprehensive radiomics features, we converted the original PET images into wavelet and Laplacian of Gaussian (LoG) images with wavelet transforms and LoG transforms, respectively. Then, three types of radiomics features were extracted from the original, wavelet and LoG images: (I) first-order statistics, (II) shape features and (III) texture features. The full list of characteristics is provided in Supplementary Table 1. The workflow of our radiomic analysis is shown in Fig. 2.

Fig. 2
figure 2

Analysis workflow in this study. SVM: support vector machine, GBDT: gradient boosting decision tree, XGBoost: extreme gradient boosting, RF: random forest, LGB: light gradient boosting machine, LASSO: least absolute shrinkage and selection operator, Logistic: logistic regression

Full size image

In addition, the conventional metabolic parameters, including the SUVmax, MTV, and TLG, were calculated by volume viewer software on a dedicated workstation (CompassView 5.0, Philips Corp., Amsterdam, the Netherlands). Regions of interest (ROIs) were manually placed to cover the lesion, and the maximal SUV (SUVmax) value was recorded for each lesion in the PET images of the patients. For each PET dataset, the SUVmax was defined as the highest SUV among all hypermetabolic tumour foci. The MTV was determined by drawing a circular ROI that included all involved lesions in the axial, coronal, and sagittal PET images. The voxel boundaries were produced automatically with the 41% SUVmax threshold method recommended by the European Association of Nuclear Medicine [28]. The total metabolic tumour volume (TMTV) was obtained by summing the MTV of all lesions. The TLG was calculated as the sum of the product of the SUVmean and MTV of all lesions (TLG = sum of [SUVmean × MTV]).

Radiomics feature selection and signature construction

First, all extracted radiomics features were harmonized using z scores. The interobserver repeatability of each radiomic feature was evaluated in 50 randomly chosen patients using the intraclass correlation coefficient (ICC) method, and features with an ICC greater than 0.75 were selected.

Then, we calculated the effective and robust R-signature based on the cross-combination method of seven machine learning models. Specifically, we choose seven classical machine learning models: support vector machines (SVMs), gradient boosting decision trees (GBDTs), extreme gradient boosting (XGBoost), random forest (RF), light gradient boosting machines (LGB), the least absolute shrinkage and selection operator (LASSO), and logistic regression (Logistic). The SVM [29] maps the input samples to points in high-dimensional space to maximize the gap between two categories. The GBDT [30] is a robust prediction model that consists of a series of weak decision trees. The XGBoost method extends GBDTs [31] in terms of structure and training strategy for more efficient predictions. Furthermore, the LGB [32] outperforms XGBoost in terms of accuracy and operation speed due to its tree structure and optimization algorithm. The RF method [33] consists of a large number of individual decision trees that operate as an ensemble. Each individual tree in the RF produces a class prediction, and the class with the most votes becomes the model’s prediction. LASSO [34] is a classical regression analysis method that performs both feature selection and regularization, with features with nonzero coefficients viewed as useful features. Logistic regression [35] combines linear regression and activation functions to estimate the probability of a certain class.

These seven models acted as feature selection and classification models for predicting the PFS and OS. The feature selection models were used to select the most significant radiomics features, and the classification models were used to construct an R-signature based on the selected features. In previous works, LASSO was generally used, as it produced a simpler and more interpretable format than the other models. However, the effectiveness of LASSO has not been thoroughly considered. In this paper, we generated a set of feature selection-classification candidates with cross-combinations of the seven machine learning models, such as LASSO-SVM, LGB-XGBoost and Logistic-RF. Then, the R-signature was built by the optimal candidate. The optimal candidate was determined with tenfold cross-validation in four steps (see Supplementary Fig. 1): (I) the samples were split into 10 groups of approximately equal size. One group was used for validation, and the remaining 9 groups were used for training. This process was repeated 10 times, with different groups used for validation each time. Considering that our samples came from two medical centres, in the split procedure, we ensured that the cohorts retained the same proportion of sample sources as the original dataset. (II) Given a training cohort, we trained a feature selection model with all PET radiomics features for predicting the PFS/OS to obtain the corresponding feature weights. Based on the feature weights, we trained a feature classification model by recursively considering the subsets of the radiomics features. (III) We identified the feature subset with the largest AUC value as containing the most important features, and the corresponding feature selection-classification model was identified as the optimal candidate. The R-signature was the prediction result of the optimal classification model. (IV) Based on the optimal candidate, we calculated the AUC of each validation cohort and outputted the average AUC of all 10 validation cohorts.

Development and validation of the models

Univariate and multivariate Cox regression analyses were used to select the potential independent prognostic factors of PFS and OS in the training cohort. All the significant clinical factors (age, LDH, B symptoms, ECOG, Ann Arbor stage, extranodal involvement, bulky disease) in the univariate analysis were included in the multivariate analysis. Independent predictors (age, Ann Abor stage, bulky disease) were used to build the clinical model. Then, all statistically significant PET factors (SUVmax, TMTV, TLG, R-signature) in the univariate analysis were included in the multivariate analysis. The independent predictors (SUVmax, TMTV, R-signature) were used to build a combined PET-based model. After that, we assembled the models by combining all the independent clinical and PET predictors to build the combined model. Calibration curves were used to assess the models. To quantify the discrimination performance of the models, the Harrell concordance index (C-index) and the time-dependent area under the ROC curve (tdROC) were calculated. The performance of the models was also externally tested in the independent validation cohort with the formula and cut-off values derived from the training cohort.

Clinical benefit analysis based on the models

To estimate the clinical utility of the models, decision curve analysis (DCA) was performed by calculating the net benefits for a range of threshold probabilities in the training cohort and validation cohort.

Statistical analysis

All statistical tests were performed using SPSS 22.0 (IBM Corp., Armonk, NY, USA) and R statistical software (version 4.0.2). A P value less than 0.05 was considered statistically significant. The difference in the related clinical information between the training and validation cohorts was assessed using χ2 tests or Mann–Whitney U tests, as appropriate. We determined the optimal threshold of the R-signature based on the receiver operating characteristic (ROC) curve. Cox regression analysis was used to analyse the potential independent predictors and build the models. The survival conditions were evaluated by the Kaplan–Meier method and compared by log-rank tests. The calibration curves, C-index, tdAUC and DCA were calculated for the models in the training and validation cohorts.

Results

Patient characteristics

The baseline characteristics of the included patients in the training and validation cohorts are summarized in Table 1. No statistically significant (P < 0.05) difference was observed between the two cohorts. A total of 14 participants had SUVmean values in the liver<1.3 (training cohort: n=3; validation cohort: n=2) or ≥3.0 (training cohort: n=5; validation cohort: n=4). The median follow-up times for the training and validation cohorts were 35.0 and 23.5 months, respectively. Of the participants who experienced disease relapse or progression in the training cohort, 103 and 76 died, respectively, and the 1-year, 3-year, and 5-year PFS and 1-year, 3-year, and 5-year OS rates were 77.3%, 66.1%, and 54.6% and 85.7%, 76.3%, and 66.8%, respectively. Of the participants who experienced disease relapse or progression in the validation cohort, 33 and 25 died, respectively, and the 1-year, 3-year, and 5-year PFS and 1-year, 3-year, and 5-year OS rates were 83.5%, 67.9%, and 63.4% and 88.9%, 74.1%, and 71.9%, respectively.

Table 1 Demographics and clinical characteristics of the study population
Full size table

Radiomics feature selection and R-signature construction

In the first step of reproducible feature selection, we excluded features with ICCs less than 0.75, and 274 features with high reproducibility were selected for further analysis. Based on the results of the feature extraction-classification models, we determined 12 radiomic features for constructing the R-signaturePFS for the SVM-SVM model (AUC=0.757) and 31 radiomic features for constructing the R-signatureOS for the XGBoost-XGBoost model (AUC=0.762) (Fig. 3 and Supplementary Tables 2 and 3). The optimal cut-off values of the R-signatures were 0.386 for PFS and 0.268 for OS. Using these cut-offs, the risk groups identified by the R-signatures showed distinct OS and PFS outcomes in both the training and validation cohorts (Fig. 3).

Fig. 3
figure 3

Heatmaps indicate the AUC performance of the cross-combinations of the feature selection methods (columns) and classification models (rows) in predicting PFS (a) and OS (c). The spectra demonstrating the selected features to build R-signaturePFS (b) and R-signatureOS (d). Kaplan–Meier plots for patients stratified by risk classification according to R-signatures for patients in the training (e, f) and validation (g, h) cohorts

Full size image

Univariate and multivariate analysis results

The differences in the clinical characteristics between the dichotomized R-signaturePFS and R-signatureOS groups are shown in Table 2.

Table 2 Comparison of the R-signature with the patient clinical data and PET metabolic parameters in the training cohort
Full size table

For the clinical variables, we performed univariate and multivariate Cox regression analyses to identify the independent factors associated with survival. We identified age (HR=2.460, P<0.001; HR=2.810, P<0.001), Ann Abor stage (HR=3.139, P<0.001; HR=3.105, P<0.001) and bulky disease (HR=1.518, P=0.043; HR=1.631, P=0.041) as independent risk factors for PFS and OS (Table 3).

Table 3 Univariate and multivariate analyses of factors predictive of progression-free survival and overall survival in the training cohort
Full size table

For the PET variables, univariate and multivariate Cox regression analyses showed that the R-signature (HR=5.024, P<0.001; HR=4.054, P<0.001), SUVmax (HR=2.027, P=0.001; HR=2.005, P=0.004) and TMTV (HR=2.265, P<0.001; HR=1.789, P=0.036) were independent risk factors for PFS and OS (Table 3).

Assessment and validation of models built for predicting PFS and OS

The independent clinical (age, Ann Abor stage, and bulky disease) and PET (R-signature, TMTV and SUVmax) predictors that were significant in the multivariate analysis were used to build the combined models for predicting the PFS and OS rates (Fig. 4 and Table 4). Individual clinical models and PET-based models were also constructed (Table 4).

Fig. 4
figure 4

a The combined model for predicting PFS. b, d Calibration curves of the model for predicting PFS in the training and validation cohorts at the 1-year, 3-year, and 5-year time points. c, e Time-dependent area under the ROC curve of the models for predicting PFS in the training and validation cohorts. f The combined model for predicting OS. g, i Calibration curves of the models for predicting OS in the training and validation cohorts at the 1-year, 3-year, and 5-year time points. h, j Time-dependent area under the ROC curve of the models for predicting OS in the training and validation cohorts

Full size image
Table 4 The survival-prediction models included in this study
Full size table

For PFS, the C-indices of the combined model in the training and validation cohorts were 0.801 (95% CI: 0.745-0.848) and 0.758 (95% CI: 0.618-0.795), respectively; these values were superior to those of the clinical model (C-index: 0.732, 95% CI: 0.679-0.785 in the training cohort, and C-index: 0.621, 95% CI: 0.512-0.731 in the validation cohort), PET-based model (C-index: 0.785, 95% CI: 0.734-0.835 in the training cohort, and C-index: 0.732, 95% CI: 0.642-0.823 in the validation cohort) and NCCN-IPI model (C-index: 0.720, 95% CI: 0.668-0.772 in the training cohort, and C-index: 0.673, 95% CI: 0.584-0.763 in the validation cohort) (Table 5).

Table 5 The Harrell’s C-index results in the training and validation cohorts
Full size table

For OS, the C-indices of the combined model in the training and validation cohorts were 0.807 (95% CI: 0.736-0.854) and 0.794 (95% CI: 0.668-0.881); these values were superior to those of the clinical model (C-index: 0.740, 95% CI: 0.680-0.800 in the training cohort, and C-index: 0.696, 95% CI: 0.584-0.807 in the validation cohort), PET-based model (C-index: 0.773, 95% CI: 0.713-0.833 in the training cohort, and C-index: 0.781, 95% CI: 0.697-0.902 in the validation cohort) and NCCN-IPI model (C-index: 0.726, 95% CI: 0.665-0.786 in the training cohort, and C-index: 0.708, 95% CI: 0.610-0.805 in the validation cohort) (Table 5).

The calibration curves for predicting the probability of PFS and OS at 1, 3, or 5 years for each model after 1000 bootstrap replicates are shown in Fig. 4. The calibration curves showed satisfactory agreement between the estimations and actual observations in the training and validation cohorts for the combined models.

Performance analysis of the combined models in clinical use

The decision curve analyses for the combined models are presented in Fig. 5. The decision curve analysis showed that the combined models had a higher overall net benefit than the clinical models, PET-based models, and NCCN-IPI across most of the risk thresholds in both the training and validation cohorts for PFS and OS.

Fig. 5
figure 5

Decision curve analysis for the models in the training (a, b) and validation (c, d) cohorts

Full size image

Outcome according to R-signature in the subgroups of the NCCN-IPI model

To confirm the added prognostic value of radiomics features to the NCCN-IPI, we evaluated the prognostic value of R-signatures in low/low-intermediate risk (NCCN-IPI of <4) and high-intermediate/high (NCCN-IPI of ≥4) groups.

For the training cohort, in the low/low-intermediate risk group (n=138), patients with a low R-signature had a better survival rate (92.3% for PFS and 89.5% for OS) than those with a high R-signature (51.1% for PFS and 35.7% for OS), and the P values were both <0.001. In the high-intermediate/high risk group (n=135), patients with a low R-signature had a better survival rate (75.7% for PFS and 77.0% for OS) than those with a high R-signature (34.7% for PFS and 39.3% for OS), and the P values were both 0.001 (Fig. 6 and Table 6).

Fig. 6
figure 6

Kaplan–Meier plots for PFS and OS in patients with different NCCN-IPI scores in relation to their R-signatures. a-d Low/low-intermediate NCCN-IPI (score: 0,1/2,3) and high-intermediate/high NCCN-IPI (score: 4,5/6-8) in relation to R-signaturePFS for PFS stratification in the training (a, b) and validation (c, d) cohorts. e-h. Low/low-intermediate NCCN-IPI (score: 0,1/2,3) and high-intermediate/high NCCN-IPI (score: 4,5/6-8) in relation to R-signatureOS for OS stratification in the training (e, f) and validation (g, h) cohorts

Full size image
Table 6 PFS and OS of patients based on their R-signatures in the low/low-intermediate risk and high-intermediate/high risk groups
Full size table

For the validation cohort, in the low/low-intermediate risk group (n=68), patients with a low R-signature had a better survival rate (97.4% for PFS and 96.2% for OS) than those with a high R-signature (55.2% for PFS and 56.3% for OS), and the P values were both statistically significant (<0.001). In the high-intermediate/high risk group (n=42), patients with a low R-signature had a better survival rate (75.0% for PFS and 83.3% for OS) than those with a high R-signature (36.4% for PFS and 33.3% for OS), and the P values were 0.031 and 0.007 (Fig. 6 and Table 6).

Discussion

Our study analysed PET radiomics features to predict survival outcomes in DLBCL patients. The results of this study indicated that combined models that integrated R-signatures calculated by radiomics features extracted from pretreatment [18F]-FDG PET using machine learning methods could be used as a preoperative tool for individualized prognosis prediction treatment guidance in patients with DLBCL.

Accurately predicting the prognosis of patients is of great importance for optimizing DLBCL treatment strategies. Several studies have attempted to assess the predictive value of radiomics information from PET images in DLBCL [21,22,23]. Aide et el. recruited 132 patients with DLBCL to evaluate the ability of pretreatment PET radiomics features to predict prognosis and found that LZLGE was the only independent predictor of 2-year event-free survival (HR=2.84, P=0.01) with ROC analysis accuracy (area under curve=0.76) [21]. Lue KH et el. extracted 80 PET-based radiomics features from 171 patients with DLBCL and found that RLNGLRLM was independently associated with PFS (HR=15.7, P=0.007) and OS (HR=8.64, P=0.040) [22]. Eertink et al. found that PET radiomics features extracted from the largest lesion of DLBCL patients were able to predict the 2-year time to progression (area under curve=0.67) [23]. Consistent with previous studies, the R-signatures consisting of PET radiomics features identified with machine learning in the present study were independently associated with PFS (HR=4.15, P<0.001) and OS (HR=4.029, P<0.001). Both our study and previous studies support the viewpoint that pivotal radiomics feature data extracted from PET images could help predict clinical outcomes in patients with DLBCL. The underlying cause might be explained by the hypothesis that radiomics features are strongly correlated with tumour heterogeneity [36,37,38], which is a prognostic determinant of patient survival [39,40,41].

Recently, machine learning models have attracted great attention in both research and industry because they can automatically detect inherent patterns in data distribution [42]. With advancements in machine learning, different machine learning models, such as random forest (RF) and support vector machines (SVMs), have been utilized to extract data features. These methods differ from each other in terms of structure, optimization algorithm, scale, and performance in different datasets. For example, SVMs are more suitable than logistic regression for large datasets [43]. However, most existing works have only focused on a single method for feature selection and R-signature construction, and few comparisons have been performed with other machine learning methods to assess robustness and efficiency. The cross-combination of feature selection-classification models has shown great efficiency and robustness in clinical practice [24, 25]. Thus, we developed a cross-combination approach for determining the optimal feature selection-classification pair based on AUC results. As shown in Fig. 3, the AUC used by the LASSO method for predicting OS was poor (0.696). Therefore, we chose the most powerful selection-classification pair from 49 possible combinations. The SVM-SVM model had an AUC of 0.757 for predicting PFS, and the XGBoost-XGBoost model had an AUC of 0.762 for predicting OS. In addition, we found that the selected features (see Fig. 3) were mainly texture features with the wavelet and LoG filtering algorithms, indicating that these features have better discrimination than shape and first-order features.

[18F]-FDG PET results can be used as a measure of cellular glucose metabolism to provide biological information about tumours. Our study demonstrated that the metabolic parameter TMTV was an independent predictor of PFS (HR=1.90, P=0.003), and SUVmax was found to be an independent predictor of both PFS (HR=2.03, P=0.001) and OS (HR=1.77, P=0.015). These results were consistent with previous studies [44,45,46,47]. To apply these results in clinical practice, we developed user-friendly predictive combined models that incorporated the R-signature, metabolic metrics (SUVmax and TMTV), and clinical risk factors (age, stage and bulky disease), and we compared the predictive performance of the models with clinical, PET-based, and NCCN-IPI prognostic models. The study found that the combined models had satisfactory agreement with the calibration curve and exhibited significant prognostic superiority over other models in terms of C-indices and tdAUCs.

To confirm the combined model’s risk discrimination, we compared the clinical utility of the combined model with that of other models (such as the clinical model, PET-based model and NCCN-IPI model). The DCA demonstrated that the combined models were superior to the other models across most of the range of reasonable threshold probabilities, thus suggesting that they are more suitable for estimating individual survival. Consistent with previous studies [6,7,8], our results showed that the NCCN-IPI might be improved for identifying high-risk groups for which novel treatment approaches are most needed (Supplementary Fig. 2 and Supplementary Table 4). The possible additional predictive value of the combined R-signature and the NCCN-IPI model was explored in this study. The results showed that the addition of the R-signature to the NCCN-IPI model yielded a better stratification for DLBCL patients in both high-intermediate/high risk (NCCN-IPI of ≥4) and low/low-intermediate risk (NCCN-IPI of <4) groups, in both the training and validation cohorts. Interestingly, several studies have shown that the TMTV also had additive prognostic value in patients in the NCCN-IPI subgroup [48, 49]. In agreement with the work of Eertink et al. [23], our results support further exploration of combining radiomics features with clinical predictors to more accurately identify high-risk patients. To uncover the potential additional prognostic value of these radiomics features, these results may contribute to better understanding of the clinically beneficial effects of these parameters with machine learning methods.

One of the limitations of our study was its retrospective nature. Additionally, the biological features of DLBCL, including the double expressor, double hit, and genomic prognostic factors, were not investigated in this study. Furthermore, the SUVmean of the liver was outside the normal range in a small portion of our patients (n=14). According to the quality control (QC) criteria described by EANM guidelines, the SUVmean of the liver should be between 1.3 and 3.0. Moreover, given the low number of progression or death events in the NCCN-IPI categories, conclusions about the additional predictive value of the R-signature to the NCCN-IPI model should be interpreted with caution, and larger cohorts are needed in the future. In addition, we used only the 3 largest lesions as targets for radiomic feature calculation in the present study. To date, there is no consensus on the tumour segmentation method for radiomic feature calculation in patients with DLBCL. Due to the high distribution variability of nodal and/or extranodal lesions with heterogeneous volumes and variable metabolic activity, lymphoma segmentation is more challenging than that of primary tumour lesions. Therefore, we proposed a concise tumour segmentation approach for the assessment of radiomics features that may be easier to apply in clinical settings. Finally, the 3D slicer and the growth cut algorithm applied in the present study have not been clinically approved. Although the development of PyRadiomics is involved in the standardization effort of the IBSI (The Image Biomarker Standardization Initiative) team, there are still some differences between PyRadiomics-based feature extraction and IBSI-recommended feature extraction. The main differences are due to grey value discretization, resampling (grid alignment, grey value rounding, and mask resampling), resegmentation, and kurtosis calculation. All of these differences may alter the extrapolation of our results to other clinical practices. The main strengths of our study are the homogeneity of the included patients (de novo DLBCL histology and R-CHOP-like regimens as standard treatment) and the use of an external validation method that supports the general applicability of the models.

Conclusion

The R-signature obtained with the cross-combination method could be used as a survival predictor for patients with DLBCL. Furthermore, the combined models that incorporate R-signature metabolic metrics and clinical risk factors may allow for accurate risk stratification and individualized management of DLBCL patients.