An interpretable PET/CT-based radiomic-clinical model for predicting bone marrow involvement in follicular lymphoma: comparison of pelvic and spine-pelvis VOI frameworks

Abstract

Purpose

To investigate the feasibility of non-invasively identifying bone marrow involvement (BMI) in follicular lymphoma (FL) using baseline ¹⁸F-FDG PET/CT combined with multidimensional feature fusion, and to compare the impact of different bone marrow volume-of-interest (VOI) frameworks on model performance.

Methods

This retrospective study included 187 patients with newly diagnosed FL, 93 of whom had BMI. Based on baseline ¹⁸F-FDG PET/CT, two bone marrow VOI frameworks were constructed: a pelvic VOI framework and a spine-pelvis VOI framework. Clinical features, conventional imaging features, radiomic features, and deep learning features were extracted. A hierarchical feature screening strategy was employed: clinical and conventional imaging features were screened using univariate logistic regression, Spearman’s correlation analysis, and multivariate logistic regression, whereas high-dimensional radiomic and deep learning features were screened using LASSO regression combined with the Boruta algorithm. Based on the selected features, six different modelling schemes were developed. The optimal scheme was selected using the area under the receiver operating characteristic curve (AUC) in the independent validation set as the primary metric. Under the optimal scheme, the performance of seven machine learning models—logistic regression (LR), support vector machine (SVM), gradient boosting machine (GBM), neural network (NN), random forest (RF), k-nearest neighbours (KNN), and adaptive boosting (AdaBoost)—was further compared. SHAP analysis was used to interpret the key features of the final model and the direction of their contributions.

Results

Compared with the non-BMI group, the BMI group was more likely to present with widespread regional lymph node involvement, B symptoms, larger lymph node lesions, as well as lower Hb, higher LDH, lower Apo A, lower eGFR, and higher β2-MG levels (all P < 0.05). Under both VOI frameworks, the BMI group exhibited higher bone marrow FDG uptake intensity and metabolic burden, as reflected by higher values of conventional PET/CT features, including SUVmean, Standard Deviation (PET), RMS, 25th Percentile Value, Median, 75th Percentile Value, TLG, Glycolysis Q2-Q4, SAM, and SUVpeak (all P < 0.05). Multivariate logistic regression analysis indicated that regional lymph node involvement and β2-MG consistently remained independent predictors across both VOI frameworks, whereas SUVmean retained statistical significance only within the pelvic VOI framework. A comparison of six modelling schemes revealed that the scheme integrating the spine-pelvis VOI framework with clinical features, conventional imaging features, and radiomic features performed best. Under this scheme, the GBM model achieved the best overall performance on the independent validation set (AUC = 0.906, Accuracy = 0.877, Precision = 0.926, Sensitivity = 0.833, Specificity = 0.926, F1 score = 0.877). SHAP analysis revealed that, in addition to LNr (≥ 5) and β2-MG, first-order statistical features such as PET-Orig-FO-IQR, as well as texture features derived from wavelet/LBP transformations—including PET-Wav-HLL-NGTDM-Strength, PET-Wav-HLL-GLRLM-SRHGLE, CT-LBP3D-m1-GLCM-MCC, and PET-LBP3D-m2-GLSZM-SAHGLE—also made significant contributions. These findings suggest that BMI-associated imaging phenotypes are characterised not only by increased bone marrow metabolism but also by remodelling of the grey-level distribution and spatial heterogeneity within the bone marrow.

Conclusion

Bone marrow involvement in follicular lymphoma is associated with higher tumour burden and altered metabolic heterogeneity within the bone marrow. A PET/CT-based radiomic-clinical model showed good performance for non-invasive BMI prediction, and the spine-pelvis VOI framework outperformed the pelvic VOI framework alone. The final GBM model may provide a feasible imaging biomarker for complementary baseline assessment of BMI in FL.

Supplementary Information

The online version contains supplementary material available at 10.1007/s00259-026-07918-y.

Introduction

Follicular lymphoma (FL) is an indolent non-Hodgkin lymphoma (NHL) originating from germinal centre B cells and characterised by marked heterogeneity in clinical course and prognosis [1]. Bone marrow involvement (BMI) is relatively common in FL, with more than 50% of patients having BMI at diagnosis [2]. As the bone marrow is one of the common sites of lymphoma cell infiltration, identifying BMI is of considerable clinical importance. Bone marrow evaluation not only helps determine the extent of lymphoma involvement but is also a crucial component of the Ann Arbor staging system; the presence of BMI defines the disease as stage IV. Although most patients with FL respond well to first-line treatment, BMI remains associated with a higher disease burden and adverse clinical outcomes and is incorporated into the initial staging assessment of FL as well as risk stratification systems such as FLIPI-2 and PRIMA-PI [3, 4]. Therefore, accurately identifying patients with concomitant BMI before treatment remains a key issue in the baseline assessment of FL, with important implications for risk stratification and treatment decisions.

Currently, assessment of BMI in patients with FL still relies primarily on bone marrow biopsy (BMB). Although BMB offers the advantage of pathological confirmation, it is invasive and can reflect only the status of a single bone marrow site; furthermore, sampling error may occur, making it difficult to comprehensively assess the extent of systemic bone marrow involvement. In contrast, ¹⁸F-FDG PET/CT provides whole-body metabolic information and has become an important imaging modality for the initial staging and response assessment of FDG-avid lymphomas [5]. However, the clinical value of PET/CT in assessing BMI varies according to lymphoma subtype. In Hodgkin lymphoma, routine BMB is no longer recommended for initial staging; in most cases of diffuse large B-cell lymphoma, additional BMB is generally considered only when PET/CT is negative and the identification of discordant bone marrow histology may influence clinical management; for other histological subtypes such as FL, current evidence remains insufficient to support completely replacing BMB with PET/CT for bone marrow assessment [6, 7]. The ESMO guidelines still recommend incorporating BMB into the initial evaluation of FL and PET/CT for routine staging [8]. Preliminary studies based on visual assessment of FDG PET/CT have suggested that it can detect BMI in patients with FL with reasonable accuracy and may serve as an independent prognostic factor [9]. However, the identification of BMI in FL may still be affected by bone marrow metabolic heterogeneity and differences in uptake patterns [10].

Radiomics offers a new research direction for the non-invasive assessment of bone marrow involvement in FL [11]. A meta-analysis in FL showed that PET/CT for detecting bone marrow involvement (BMI) had a pooled sensitivity of 0.67, a specificity of 0.82, and an AUC of 0.83, suggesting that PET/CT remains insufficient to replace bone marrow biopsy (BMB) on its own [12]. Compared with visual assessment based on focal abnormal uptake and limited semiquantitative indicators such as SUV, radiomics holds promise for identifying differences in bone marrow metabolic distribution and spatial heterogeneity. A small retrospective study that developed a predictive model based on skeletal texture features derived from baseline ¹⁸F-FDG PET/CT reported a significant difference between bone marrow biopsy and visual PET assessment (P = 0.010), but no significant difference between bone marrow biopsy and the PET predictive score (P = 0.097) [13]. Currently, quantitative imaging studies in FL remain limited and are predominantly retrospective; differences in segmentation methods, preprocessing workflows, and model validation strategies across studies also limit the reproducibility of findings and their clinical translation. Furthermore, studies that systematically model bone marrow VOIs and integrate clinical information with multidimensional imaging features to predict BMI are still lacking, and the impact of different bone marrow VOI frameworks on BMI identification remains unclear.

This study aimed to develop a method for predicting BMI based on multidimensional feature fusion from baseline ¹⁸F-FDG PET/CT and to compare model performance across different bone marrow VOI strategies. Using baseline ¹⁸F-FDG PET/CT images from patients with newly diagnosed FL, we systematically evaluated the value of combining clinical features with conventional imaging features, radiomic features, and deep learning features for BMI identification across different bone marrow VOI frameworks, and compared the performance of different modelling strategies. The study aimed to establish a non-invasive assessment strategy that balances stability, interpretability, and clinical feasibility, thereby providing supplementary information for pre-treatment BMI identification and complementing the initial staging and risk assessment of FL.

Methods

Study population

This single-centre retrospective cohort study included 187 patients with histopathologically confirmed FL treated at Nanjing Drum Tower Hospital, the Affiliated Hospital of Nanjing University Medical School, between January 2013 and July 2025. All patients underwent baseline ¹⁸F-FDG PET/CT and bone marrow biopsy (BMB) at initial diagnosis and had received no prior anti-lymphoma therapy. Patients were randomly assigned to the training and validation sets in a 7:3 ratio. The cohort included 84 men and 103 women aged 18–85 years (mean age, 54.07 ± 12.35 years). This study was conducted in accordance with the Declaration of Helsinki and was approved by the Ethics Committee of Nanjing Drum Tower Hospital, the Affiliated Hospital of Nanjing University Medical School (approval No. 2024-851-01). The requirement for informed consent was waived. The inclusion criteria were as follows: (1) biopsy-confirmed histopathological diagnosis of follicular lymphoma; (2) an interval of no more than 4 weeks between baseline PET/CT and bone marrow biopsy (BMB); and (3) age ≥ 18 years. The exclusion criteria were as follows: (1) receipt of lymphoma-directed or other anticancer treatment before PET/CT; (2) incomplete clinical data or loss to follow-up; and (3) concomitant other malignant tumours.

¹⁸F-FDG PET/CT imaging acquisition

¹⁸F-FDG PET/CT imaging was performed using the Philips GEMINI GXL 16 PET/CT and Philips Vereos PET/CT scanners. Intravenous injection of ¹⁸F-FDG (manufactured by Nanjing Jiangyuan Andico Positron Research and Development Co., Ltd., radiochemical purity > 95%) at a dose of 3.7 MBq/kg was administered based on patient body weight. Fasting blood glucose levels were measured prior to injection. Patients fasted for at least 6 h prior to the examination, maintaining a fasting blood glucose level of ≤ 11.1 mmol/L. After radiotracer injection, patients rested for 50–60 min. Approximately 500 ml of water was ingested before scanning to fill the stomach cavity. Scanning was performed in the supine position with calm breathing. Patients were positioned supine in the centre of the bed with arms raised alongside the head. The scanning range extended from the cranial vault to the mid-femur. CT acquisition parameters: tube voltage 120 kV, tube current 100 mA, slice thickness 2.0 mm. PET acquisition comprised 7–10 bed positions, each scanned for 1 min, matrix size 144 × 144. CT data were used for attenuation correction. Corrected PET images were reconstructed using the ordered subset expectation maximization (OSEM) method.

Definition of the outcome variable

The diagnosis of BMI was based on BMB results and visual assessment of ¹⁸F-FDG PET/CT images. Patients with bone marrow involvement confirmed by baseline BMB were directly classified as BMI-positive. For patients whose baseline BMB showed no evidence of bone marrow involvement, further visual assessment of PET/CT images was performed. If focally increased ¹⁸F-FDG uptake was observed in the bone marrow of the spine or pelvis, and this increased uptake could not be explained by the corresponding CT findings or clinical history, the patient was classified as BMI-positive. Normal bone marrow uptake was defined as uptake not exceeding that of the liver. When baseline PET/CT showed diffusely increased ¹⁸F-FDG uptake in the bone marrow (i.e., higher than liver uptake), an integrated assessment was performed based on the results of interim PET/CT and interim BMB after systemic therapy. If interim PET/CT showed reduced bone marrow uptake compared with baseline, and interim BMB confirmed bone marrow involvement, the patient was classified as BMI-positive. All other cases were classified as BMI-negative.

PET/CT images were independently reviewed by two nuclear medicine physicians with more than 10 and 15 years of PET/CT interpretive experience, respectively; in cases of disagreement, a senior physician with more than 20 years of image interpretation experience performed a further review and made the final determination.

¹⁸F-FDG PET/CT radiomic analysis with machine learning

The analysis workflow consisted of three stages: (1) constructing pelvic and spine-pelvis VOIs on CT images and extracting conventional imaging features, radiomic features, and deep learning features from the registered PET and CT images within the corresponding frameworks; (2) screening different types of features to identify candidate variables; and (3) establishing machine learning models based on different VOI frameworks and feature combinations and comparing their predictive performance for BMI (Fig. 1).

Fig. 1.

Overview of the study design. (A) A total of 187 patients with newly diagnosed follicular lymphoma were retrospectively enrolled and divided into a training set (n = 130) and a validation set (n = 57). Bone marrow assessment results and visual assessment of ¹⁸F-FDG PET/CT images were used to determine the bone marrow involvement status of the patients. (B) Based on pelvic and spine-pelvis VOI segmentation, conventional imaging features, radiomic features, and deep learning features were extracted from PET/CT images. (C) A BMI prediction model was developed through feature selection and multimodal feature integration

Image feature extraction

This study used the MONAI Auto3DSeg plugin in 3D Slicer version 5.10.0 to automatically segment CT images using the pretrained “Hip and Spine (v1.2.0)” model [14]. The left hip, right hip, and sacrum were combined to define the “pelvis” VOI, whereas the C1-C7, T1-T12, and L1-L5 vertebrae were combined to define the “spine” VOI. After automatic segmentation, all VOIs were manually reviewed and corrected to remove non-target bone regions and irrelevant structures. Subsequently, the final CT VOIs were overlaid on the fused PET images for further review to exclude non-osseous areas of abnormal uptake, such as extraosseous lesions adjacent to bone and high urinary bladder uptake. In addition, regions with contiguous bone involvement, osteophytes, and sclerotic bone changes were manually excluded, ultimately yielding the “pelvis” VOI and “spine-pelvis” VOI for feature extraction.

Within the final VOI framework, conventional imaging features were extracted using MMIS software. Conventional CT features included first-order statistics of VOI voxel count, volume, and CT attenuation values, whereas conventional PET features included first-order statistics of uptake intensity and metabolism-related parameters, such as mean, maximum, RMS, metabolic volume, and TLG. Radiomic features were also computed using MMIS software; their definitions and nomenclature followed the standards of the Imaging Biomarker Standardisation Initiative (IBSI), and software information, together with key image-processing and feature-calculation parameters, was documented in accordance with IBSI reporting recommendations [15]. Both PET and CT images were discretised using a fixed bin number (FBN) strategy and uniformly quantised into 64 grey levels [15]. In addition to the original images, PET and CT images underwent wavelet transformation and three-dimensional local binary pattern (LBP-3D) filtering to generate derived images; subsequently, first-order statistical and texture features (GLCM, GLRLM, GLSZM, GLDM, and NGTDM) were extracted from both the original and derived images, and morphological features were extracted from the original images [16–18]. A total of 2,260 radiomic features were extracted, including 1,130 PET features and 1,130 CT features.

Deep learning features were also extracted using MMIS software. A ResNet50-based convolutional neural network was used as the feature encoder to automatically identify the most representative slice within the 3D VOI, defined as the slice with the largest ROI coverage, and the ROI was expanded by 10% at the margins to preserve information from adjacent tissues. Deep learning features were extracted from the penultimate layer of the network, yielding a 2,048-dimensional feature vector from each of the PET and CT images for every case; for fusion analysis, these vectors were concatenated to form a 4,096-dimensional joint deep feature representation.

Feature selection

This study incorporated clinical features, conventional imaging features, high-dimensional radiomic features, and deep learning features and employed a hierarchical screening strategy to reduce dimensionality, minimise redundancy, and improve model robustness. For clinical and conventional imaging features, between-group comparisons were first performed on the full cohort according to BMI status to characterise the clinical profile of BMI-positive versus BMI-negative patients. Continuous variables were analysed using the Wilcoxon rank-sum test, while categorical variables were analysed using Fisher’s exact test; variables with P < 0.05 were included in the candidate set. It should be noted that this initial descriptive screening step was conducted prior to the training-validation split; all subsequent modelling steps, including univariate logistic regression, Spearman’s correlation analysis, multivariate logistic regression, LASSO regression, and the Boruta algorithm, were performed exclusively within the training set, and variables with P < 0.05 were retained. To reduce information redundancy and multicollinearity, Spearman’s correlation analysis was performed on the candidate variables before multivariate analysis; when the absolute value of the correlation coefficient was ≥ 0.80, the variables were considered strongly correlated, and redundant variables were removed. The variables retained after redundancy removal were entered into a multivariate logistic regression model. The Firth penalised likelihood method was used preferentially, together with backward elimination for stepwise variable selection; if the Firth method was not applicable, standard logistic regression was used, and variable selection was based on the likelihood ratio test, with independent predictors with P < 0.05 ultimately retained. For high-dimensional radiomic features and deep learning features, initial screening was performed using LASSO-penalised logistic regression in the training set. The optimal penalty parameter was determined by 10-fold cross-validation, and features with non-zero coefficients were retained according to the λ1se criterion. Subsequently, secondary screening was performed using the Boruta algorithm, with feature importance iteratively evaluated based on a random forest model. Using shadow features as controls, features classified as Confirmed were ultimately retained for model construction.

Model development and validation

Based on two VOI frameworks (pelvis and spine-pelvis) and three feature combinations (radiomics; clinical-conventional imaging-radiomics; and clinical-conventional imaging-radiomics-deep learning), a total of six modelling schemes were designed: Scheme A (pelvis + radiomics), Scheme B (pelvis + clinical-conventional imaging-radiomics), Scheme C (pelvis + clinical-conventional imaging-radiomics-deep learning), Scheme D (spine-pelvis + radiomics), Scheme E (spine-pelvis + clinical-conventional imaging-radiomics), and Scheme F (spine-pelvis + clinical-conventional imaging-radiomics-deep learning) (Fig. 2). Under each modelling scheme, seven machine learning algorithms were used to build models, including logistic regression (LR), support vector machine (SVM), gradient boosting machine (GBM), neural network (NN), random forest (RF), k-nearest neighbours (KNN), and adaptive boosting (AdaBoost).

Fig. 2.

Schematic illustration of the six modelling schemes: pelvis + radiomics (A), pelvis + clinical-conventional imaging-radiomics (B), pelvis + clinical-conventional imaging-radiomics-deep learning (C), spine-pelvis + radiomics (D), spine-pelvis + clinical-conventional imaging-radiomics (E), and spine-pelvis + clinical-conventional imaging-radiomics-deep learning (F)

For each modelling scheme and algorithm combination, repeated 10-fold cross-validation with five repeats was performed in the training set for model training and hyperparameter tuning, with the mean AUC used as the optimisation metric to determine the optimal hyperparameters. All preprocessing steps were completed within each resampled training fold. Subsequently, the final model was fitted to the entire training set using the optimal hyperparameters, and its generalisation performance was evaluated on the independent validation set. Model performance was assessed using receiver operating characteristic (ROC) curves and the area under the ROC curve (AUC). The optimal classification threshold was determined using the Youden index, and accuracy, precision (equivalent to positive predictive value), sensitivity, specificity, and F1 score were calculated at that threshold. Selection of the optimal model and modelling scheme was based primarily on the AUC in the independent validation set; when differences in AUC were small, F1 score, calibration curves, and decision curve analysis (DCA) results in the validation set were further considered for comprehensive evaluation to determine the optimal modelling scheme and corresponding algorithm. For the final model, interpretability analysis was further performed using the SHAP method.

Statistical analysis

Continuous variables were expressed as median (interquartile range) or mean ± standard deviation, whereas categorical variables were presented as counts and percentages. For between-group comparisons, the Wilcoxon rank-sum test was used for continuous variables, and Fisher’s exact test was used for categorical variables. Logistic regression results were presented as odds ratios (ORs) with 95% confidence intervals (CIs). All data processing, model development, and statistical analyses were performed using R version 4.5.1 and Python version 3.12.0. All tests were two-sided, and P < 0.05 was considered statistically significant.

Results

Baseline comparison of clinical and PET/CT characteristics between BMI and non-BMI FL patients

This study included 187 patients with FL, of whom 93 were BMI-positive. Compared with the non-BMI group, the BMI group showed significant differences in several variables, including regional lymph node involvement (88.2% vs. 41.5%, P < 0.001), B symptoms (38.7% vs. 6.4%, P < 0.001), maximum lymph node diameter > 6 cm (33.3% vs. 17.2%, P = 0.018), haemoglobin (Hb) level [123.00 (113.50, 138.00) vs. 131.00 (123.00, 143.00), P = 0.005], lactate dehydrogenase (LDH) level [192.00 (168.00, 246.50) vs. 178.00 (155.00, 213.00), P = 0.04], apolipoprotein A (Apo A) concentration [0.96 (0.84, 1.11) vs. 1.06 (0.92, 1.27), P < 0.001], estimated glomerular filtration rate (eGFR) [103.25 (88.50, 115.30) vs. 109.60 (96.50, 129.10), P = 0.038], and β2-microglobulin (β2-MG) level [2,390.00 (1,988.00, 3,462.00) vs. 1,848.50 (1,525.00, 2,246.00), P < 0.001] (Table 1). Furthermore, significant differences in multiple conventional imaging features were observed between the BMI and non-BMI groups under both the pelvic VOI and spine-pelvis VOI frameworks. These differences mainly involved indices reflecting bone marrow FDG uptake intensity and metabolic burden, including SUVmean, Standard Deviation (PET), RMS, 25th percentile value, median, 75th percentile value, TLG, glycolysis Q2-Q4, SAM, and SUVpeak (all P < 0.05; Tables S1-S2).

Table 1.

Comparison of baseline clinical characteristics between BMI and non-BMI patients with follicular lymphoma

Variables	Total (n = 187)	Non-BMI (n = 94)	BMI (n = 93)	P
Age	55.00 (46.00, 63.00)	54.50 (45.00, 64.00)	55.00 (46.00, 61.00)	0.464
Sex				0.379
Female	103 (55.1)	55 (58.5)	48 (51.6)
Male	84 (44.9)	39 (41.5)	45 (48.4)
LNr (≥ 5)				< 0.001*
No	66 (35.3)	55 (58.5)	11 (11.8)
Yes	121 (64.7)	39 (41.5)	82 (88.2)
B symptoms				< 0.001*
No	145 (77.5)	88 (93.6)	57 (61.3)
Yes	42 (22.5)	6 (6.4)	36 (38.7)
LNmax (> 6 cm)	n = 186	n = 93	n = 93	0.018*
No	139 (74.7)	77 (82.8)	62 (66.7)
Yes	47 (25.3)	16 (17.2)	31 (33.3)
Histologic grade	n = 185	n = 93	n = 92	0.225
1–2	156 (84.3)	75 (80.6)	81 (88.0)
3a	29 (15.7)	18 (19.4)	11 (12.0)
Ki67 (%)	n = 179	n = 86	n = 93	0.221
	20.00 (10.00, 30.00)	20.00 (15.00, 30.00)	20.00 (10.00, 30.00)
Hb (g/L)	n = 169	n = 81	n = 88	0.005*
	128.00 (119.00, 140.00)	131.00 (123.00, 143.00)	123.00 (113.50, 138.00)
PLT (×109/L)	n = 167	n = 81	n = 86	0.103
	183.00 (143.00, 224.00)	192.00 (161.00, 224.00)	177.50 (131.00, 224.00)
ALC/AMC	n = 165	n = 79	n = 86	0.787
	3.40 (2.33, 5.00)	3.40 (2.50, 4.60)	3.37 (2.33, 5.67)
LDH (U/L)	n = 161	n = 77	n = 84	0.04*
	187.00 (161.00, 227.00)	178.00 (155.00, 213.00)	192.00 (168.00, 246.50)
Apo A (g/L)	n = 161	n = 77	n = 84	< 0.001*
	1.00 (0.89, 1.15)	1.06 (0.92, 1.27)	0.96 (0.84, 1.11)
Apo B (g/L)	n = 161	n = 77	n = 84	0.297
	0.75 (0.64, 0.91)	0.74 (0.59, 0.89)	0.78 (0.67, 0.92)
eGFR (MDRD)	n = 143	n = 65	n = 78	0.038*
	105.50 (90.40, 123.80)	109.60 (96.50, 129.10)	103.25 (88.50, 115.30)
β2-MG (ng/mL)	n = 124	n = 58	n = 66	< 0.001*
	2,134.00 (1,727.50, 2,732.00)	1,848.50 (1,525.00, 2,246.00)	2,390.00 (1,988.00, 3,462.00)

Variables marked with an asterisk (*) are considered statistically significant (*P < 0.05)

Feature selection

To further identify clinical and conventional imaging factors associated with bone marrow involvement in FL, univariate logistic regression was first performed to screen variables that were statistically significant in the baseline comparisons. Univariate logistic regression based on clinical features and conventional imaging features within the pelvic VOI framework showed that regional lymph node involvement, B symptoms, β2-microglobulin (β2-MG) level, apolipoprotein A (Apo A) concentration, haemoglobin (Hb) level, SUVmean, RMS, 25th percentile value, median, 75th percentile value, upper adjacent value (UAV), TLG (g), glycolysis Q2 (g), glycolysis Q3 (g), SAM (g), and SUVpeak were significantly associated with BMI (P < 0.05, Table S3). Univariate logistic regression based on clinical features and conventional imaging features within the spine-pelvis VOI framework showed that regional lymph node involvement, B symptoms, β2-microglobulin (β2-MG) level, apolipoprotein A (Apo A) concentration, haemoglobin (Hb) level, SUVmean, Standard Deviation (PET), RMS, 25th percentile value, median, 75th percentile value, upper adjacent value (UAV), TLG (g), glycolysis Q2 (g), glycolysis Q3 (g), SAM (g), SAM background, and SUVpeak were significantly associated with BMI (P < 0.05, Table S4). Candidate variables identified by univariate logistic regression were entered into Spearman’s correlation analysis to assess inter-variable correlations and reduce the risk of multicollinearity. Within the pelvic VOI framework, 14 pairs of strongly correlated variables were identified. For each pair, only one variable was retained, with priority given to the variable showing a stronger association with BMI and better clinical interpretability, resulting in 11 relatively independent candidate variables (Fig S1). Within the spine-pelvis VOI framework, 21 pairs of strongly correlated variables were identified, and 11 candidate variables were retained after redundancy removal (Fig S2). After the candidate variables identified by Spearman’s correlation analysis were entered into multivariate logistic regression, the model based on clinical features and conventional imaging features within the pelvic VOI framework showed that regional lymph node involvement [OR 3.916, 95% CI: 1.172–13.089; P = 0.027], β2-microglobulin (β2-MG) level [OR 1.001, 95% CI: 1.000-1.001; P = 0.048], and SUVmean [OR 4.573, 95% CI: 1.010–20.710; P = 0.049] remained independent risk factors (Table S3). The model based on clinical features and conventional imaging features within the spine-pelvis VOI framework showed that regional lymph node involvement [OR 4.697, 95% CI: 1.429–15.444; P = 0.011] and β2-microglobulin (β2-MG) level [OR 1.001, 95% CI: 1.000-1.001; P = 0.020] remained independent risk factors (Table S4).

A total of 2,260 radiomic features and 4,096 ResNet50-based deep learning features were extracted from each patient’s PET/CT images. Feature dimensionality reduction was performed using LASSO regression combined with the Boruta algorithm to identify the subset of features with the best discriminatory ability for BMI. In the radiomic feature analysis, 8 and 9 independent predictive features were retained under the pelvic VOI and spine-pelvis VOI frameworks, respectively (Figs S3 and S4). In the fusion analysis incorporating both radiomic and deep learning features, 8 key features (including 5 radiomic features and 3 deep learning features) and 7 key features (including 5 radiomic features and 2 deep learning features) were selected under the pelvic VOI and spine-pelvis VOI frameworks, respectively, for model construction (Figs S3 and S4).

Construction and evaluation of machine learning models

Model performance comparison

Based on ROC analysis in the independent validation set, which compared the discriminatory performance of the six modelling schemes (Fig. 3), Scheme E was ultimately identified as the optimal modelling scheme, and the performance of different machine learning algorithms was further compared under this scheme.

Fig. 3.

Comparison of ROC curves of the six predefined modelling schemes for predicting bone marrow involvement (BMI) in follicular lymphoma in the independent validation set: pelvis + radiomics (A), pelvis + clinical-conventional imaging-radiomics (B), pelvis + clinical-conventional imaging-radiomics-deep learning (C), spine-pelvis + radiomics (D), spine-pelvis + clinical-conventional imaging-radiomics (E), and spine-pelvis + clinical-conventional imaging-radiomics-deep learning (F)

After identifying the optimal modelling scheme (Scheme E), we further compared the predictive performance of seven machine learning (ML) models under this scheme in the training set (n = 130) and the independent validation set (n = 57): logistic regression (LR), support vector machine (SVM), gradient boosting machine (GBM), neural network (NN), random forest (RF), k-nearest neighbours (KNN), and adaptive boosting (AdaBoost).

In the training set (Table 2; Fig. 4A), all models showed good discrimination in the training set; among them, the GBM model showed the most balanced performance (AUC = 0.967, Accuracy = 0.938, Precision = 0.923, Sensitivity = 0.952, Specificity = 0.925, F1 score = 0.938). It should be noted that the RF and AdaBoost models demonstrated near-perfect classification results in the training set, suggesting a risk of overfitting or overly optimistic performance estimates.

Table 2.

Performance comparison of different machine learning models in the training set

Model	AUC	Accuracy	Precision	Sensitivity	Specificity	F1 Score
LR	0.912	0.854	0.844	0.857	0.851	0.850
SVM	0.927	0.869	0.883	0.841	0.896	0.862
GBM	0.967	0.938	0.923	0.952	0.925	0.938
NN	0.954	0.908	0.905	0.905	0.910	0.905
RF	1	1	1	1	1	1
KNN	0.990	0.946	0.938	0.952	0.940	0.945
AdaBoost	1	1	1	1	1	1

Fig. 4.

Performance comparison of different machine learning models under the optimal modelling scheme (Scheme E): ROC curves for the training set (A) and the independent validation set (B)

In the independent validation set (Table 3; Fig. 4B), the GBM model maintained relatively stable generalisation performance (AUC = 0.906, Accuracy = 0.877, Precision = 0.926, Sensitivity = 0.833, Specificity = 0.926, F1 score = 0.877), with the highest AUC, Accuracy, Precision, Specificity, and F1 score among all candidate models. Compared with the AdaBoost model, which is also an ensemble learning model (validation set AUC = 0.901, Accuracy = 0.842, Precision = 0.818, Sensitivity = 0.900, Specificity = 0.778, F1 score = 0.857), the GBM model performed better in terms of overall discrimination and classification consistency (F1 score). Furthermore, both the RF and AdaBoost models exhibited varying degrees of performance decline in the validation set (RF: AUC = 0.879, F1 = 0.833; AdaBoost: AUC = 0.901, F1 score = 0.857), raising the possibility of overfitting in both models.

Table 3.

Performance comparison of different machine learning models in the validation set

Model	AUC	Accuracy	Precision	Sensitivity	Specificity	F1 Score
LR	0.888	0.825	0.794	0.900	0.741	0.844
SVM	0.886	0.807	0.852	0.767	0.852	0.807
GBM	0.906	0.877	0.926	0.833	0.926	0.877
NN	0.842	0.807	0.757	0.933	0.667	0.836
RF	0.879	0.825	0.833	0.833	0.815	0.833
KNN	0.879	0.825	0.917	0.733	0.926	0.815
AdaBoost	0.901	0.842	0.818	0.900	0.778	0.857

The KNN model achieved an AUC of 0.990 and an F1 score of 0.945 in the training set, but in the validation set, the AUC dropped to 0.879, the F1 score fell to 0.815, and sensitivity decreased to 0.733. Although it maintained high precision and specificity in the validation set (Precision = 0.917, Specificity = 0.926), its limited sensitivity resulted in overall performance inferior to that of the GBM model. The NN model exhibited a “high sensitivity-low specificity” pattern in the validation set (Sensitivity = 0.933, Specificity = 0.667, F1 score = 0.836), indicating an increased false-positive rate. The LR model showed relatively stable performance in the training set (AUC = 0.912, F1 score = 0.850) and the validation set (AUC = 0.888, F1 score = 0.844), but its specificity and precision in the validation set were relatively low (Specificity = 0.741, Precision = 0.794), resulting in overall performance inferior to that of the GBM model. The SVM model maintained high specificity in both the training set (AUC = 0.927, Specificity = 0.896) and the validation set (AUC = 0.886, Specificity = 0.852), but its sensitivity in the validation set was relatively low (Sensitivity = 0.767), resulting in an F1 score (0.807) lower than that of the GBM model.

Selection of the best-performing model

After a comprehensive evaluation of the overall performance of all candidate models across all datasets, we ultimately selected the GBM model as the optimal model for this study (Fig. 5) for the following reasons:

Fig. 5.

Radar plots of the final model in the training set (A) and the independent validation set (B)

Consistent performance: The GBM model maintained high discriminative performance in both the training and validation sets, with only limited performance attenuation across datasets (training set AUC = 0.967, F1 score = 0.938; validation set AUC = 0.906, F1 score = 0.877). Robust discriminative power: In the validation set, the GBM model ranked highest across overall performance metrics, including AUC, accuracy, precision, specificity, and F1 score, thereby avoiding the limitations seen in models with either high sensitivity but insufficient specificity or high specificity but reduced sensitivity. Calibration reliability: Calibration curves showed good agreement between the GBM model’s predicted probabilities and the observed probabilities (Fig. 6A). Clinical benefit: Decision curve analysis (DCA) indicated that the GBM model provided a relatively stable net benefit across a clinically meaningful range of threshold probabilities (Fig. 6B). Interpretability and clinical utility: SHAP quantified the contributions of key variables to individual predictions and ranked them by importance, thereby providing a basis for clinical interpretation and mechanistic discussion (Fig. 6C). Therefore, the GBM model was adopted as the final model for subsequent analyses and presentation.

Fig. 6.

Calibration curve (A), decision curve analysis (B), and SHAP interpretability analysis of the final model (C)

Model interpretation

As shown in Fig. 6C, to explain the basis of the final GBM model’s predictions, this study used Shapley additive explanations (SHAP) to quantify the marginal contribution of each feature to the predicted probability of BMI and ranked feature importance according to the mean absolute SHAP value (mean |SHAP|).

Global importance analysis showed that LNr (≥ 5) contributed the most (mean |SHAP| = 0.712, accounting for 24.71%), followed by PET-Wav-HLL-NGTDM-Strength (0.322, 11.16%), PET-Orig-FO-IQR (0.279, 9.69%), PET-Wav-HLL-GLRLM-SRHGLE (0.247, 8.55%), CT-LBP3D-m1-GLCM-MCC (0.237, 8.24%), and PET-LBP3D-m2-GLSZM-SAHGLE (0.220, 7.63%). The remaining features, in descending order of importance, were PET-Wav-HLL-FO-P90 (0.202; 6.99%), PET-Wav-LLL-GLSZM-GLNN (0.193; 6.68%), PET-Orig-FO-MAD (0.184; 6.39%), and PET-Wav-HLL-GLDM-SDHGLE (0.152; 5.27%); the clinical variable β2-MG also contributed to the model (0.135, 4.69%).

SHAP direction analysis showed that when LNr (≥ 5) took high values, its SHAP values were mainly distributed in the positive region, suggesting that higher LNr was associated with an increased probability of BMI positivity. For continuous variables, higher values of PET-Orig-FO-IQR, PET-Wav-HLL-GLRLM-SRHGLE, PET-Wav-HLL-FO-P90, PET-Orig-FO-MAD, and PET-Wav-HLL-GLDM-SDHGLE generally corresponded to positive SHAP values, whereas higher values of PET-Wav-HLL-NGTDM-Strength, CT-LBP3D-m1-GLCM-MCC, PET-LBP3D-m2-GLSZM-SAHGLE, and PET-Wav-LLL-GLSZM-GLNN were more often associated with negative SHAP values. Elevated β2-MG generally showed a positive contribution trend. In summary, the model’s key information sources mainly included a clinical indicator reflecting disease burden (LNr) and β2-MG, while also relying on various texture and first-order statistical features derived from wavelet and LBP transformations to characterize the grey-level distribution and spatial heterogeneity within the bone marrow.

Discussion

In the initial evaluation of FL, BMI has clear clinical importance because it not only affects Ann Arbor staging but is also incorporated into risk stratification systems such as FLIPI, FLIPI-2, and PRIMA-PI, thereby directly influencing prognostic assessment and treatment decisions. Previous studies have shown that bone marrow biopsy remains the traditional standard for assessing BMI; however, its invasiveness, sampling limitations, and limited ability to reflect systemic bone marrow burden restrict its role in precise risk stratification. ¹⁸F-FDG PET/CT can provide whole-body metabolic information; however, current imaging assessment of bone marrow involvement in FL still relies mainly on visual assessment or limited single-parameter analysis, and systematic studies of bone marrow heterogeneity phenotypes and VOI definition strategies remain insufficient. This study found that, compared with the pelvic VOI alone, the combined spine-pelvis VOI framework achieved superior predictive performance after integrating clinical, conventional imaging, and radiomic features, suggesting that broader characterisation of bone marrow phenotypes may improve the non-invasive identification of BMI. SHAP analysis also revealed that model predictions did not rely solely on bone marrow uptake intensity itself, but rather on the combined effects of disease burden indicators and higher-order texture features; the imaging phenotype of BMI was more likely to reflect remodelling of spatial heterogeneity within the bone marrow rather than a simple increase in metabolism.

Although existing studies on imaging assessment of bone marrow involvement in lymphoma have suggested that radiomics has greater discriminatory potential than conventional PET metrics, methodological comparisons specific to FL, particularly those focusing on the extent of bone marrow volume-of-interest (VOI) coverage, remain limited. A previous retrospective study of 97 patients explored the value of [18F]FDG-PET/CT radiomics in predicting bone marrow involvement (BMI) in mantle cell lymphoma. The results showed that [18F]FDG-PET radiomic features demonstrated better classification performance than SUV alone (AUC: 0.82 vs. 0.68), and that predictive performance generally improved with increasing relative or absolute percentages of bone marrow involvement (REL and ABS) [19]. Another two-centre study further explored the feasibility of interpretable machine learning models for predicting BMI in lymphoma; however, its region-of-interest (ROI) coverage was mainly limited to the pelvic region and it did not focus on FL, an indolent lymphoma subtype with relatively distinctive biological behaviour [20]. An important advance of the present study was the direct comparison of two bone marrow phenotyping frameworks, namely the pelvic VOI and spine-pelvis VOI, and confirmation that VOI definition itself significantly affected feature selection results and model performance. This finding has clear anatomical and biological plausibility: active bone marrow in adults is primarily distributed in the axial skeleton. Previous quantitative studies based on FLT-PET have shown that the pelvis accounts for approximately 25.3% of total-body proliferative bone marrow in adults, whereas the combined pelvis and spine account for approximately 75.3%, suggesting that the spine-pelvis VOI more closely approximates the principal distribution of active bone marrow throughout the body than the pelvic VOI alone [21]. In an analysis of 261 patients with FL, Nakajima et al. identified a total of 780 focal bone marrow lesions among 78 patients with PET-defined bone marrow involvement; the vertebrae were the most frequently affected anatomical structure (33.3%), and notably, 91.2% of all PET-positive lesions were localised outside the iliac crests [9]. Furthermore, Mattonen et al. demonstrated that bone marrow radiomic features extracted from the L3-L5 vertebral bodies on baseline ¹⁸F-FDG PET/CT provided independent prognostic value for disease-free survival in patients with non-small cell lung cancer, suggesting that the clinical information captured by vertebral bone marrow radiomics may be transferable across different oncological contexts [22]. Therefore, in imaging-based modelling of FL bone marrow involvement, expanding VOI coverage to include spinal bone marrow information may more faithfully characterise the overall burden and spatial heterogeneity of bone marrow involvement. This also suggests that PET/CT assessment of bone marrow involvement in FL should not be limited to visual diffuse uptake patterns or a single SUV metric, but should further focus on grey-level distribution, local texture structure, and the bone marrow heterogeneity reflected by these features.

The findings of this study further support the notion that BMI is not an isolated local pathological phenomenon, but rather more likely represents a state of increased systemic disease burden and greater biological activity. Compared with patients without BMI, those with BMI more frequently presented with widespread regional lymph node involvement, B symptoms, larger lymph node lesions, and abnormalities such as decreased haemoglobin (Hb) and elevated lactate dehydrogenase (LDH) and β2-microglobulin (β2-MG) levels. These differences collectively point to a higher tumour burden and poorer host status. In the final model, the most important variables did not derive exclusively from high-dimensional radiomic features, but instead included the clinical burden indicator LNr, the serological marker β2-MG, and various texture features. This suggests that the key to BMI prediction lies in the integration of clinical information and imaging heterogeneity, rather than in any single indicator. The prognostic value of LNr has also been validated in previous large-sample studies of FL [23]. The joint significance of β2-MG and BMI has also been confirmed in previous studies, which is consistent with the main predictive clinical variables identified in this study [3]. In addition, the inclusion of both β2-MG and bone marrow involvement in the simplified PRIMA-PI prognostic scoring system further underscores the synergistic role of β2-MG and BMI in FL risk stratification. Moreover, elevated β2-MG levels are often positively associated with adverse factors such as elevated LDH, older age, and a greater number of involved lymph node regions, and they increase with higher FLIPI risk categories, suggesting that they may reflect broader disease dissemination and greater biological activity [24]. The results of this study not only support the close association between BMI and systemic disease burden, but also suggest that metabolic heterogeneity within the bone marrow may represent an important PET/CT imaging phenotype of this high-burden state. Although the present study focused on predicting the presence or absence of BMI, the quantitative relationships between β2-MG and the degree of BMI, extent of infiltration, and even changes in the bone marrow microenvironment still warrant further clarification in future studies.

At the model level, this study compared the predictive performance of various machine learning algorithms within the optimal VOI framework. The results showed that the GBM model achieved relatively balanced and stable performance in the independent validation set, suggesting that it had better generalisation ability under the current sample size and feature structure. Compared with RF and AdaBoost, which performed nearly perfectly in the training set but showed a marked decline in the validation set, GBM better balanced fitting ability and generalisation ability. Although LR and SVM were generally stable, their overall performance in terms of sensitivity and specificity still lagged behind that of GBM. These results suggest that for the complex task of predicting bone marrow involvement in FL, which involves clinical variables, conventional PET parameters, and high-dimensional radiomic features, simply increasing model complexity does not necessarily lead to better generalisation performance. The fact that the optimal modelling framework identified in this study did not retain deep learning features also has methodological implications. This does not imply that deep learning features lack potential; rather, it suggests that under conditions of a relatively limited sample size, increased feature complexity does not necessarily translate into more stable generalisation performance. Existing systematic reviews have noted that although deep learning models and fusion models can achieve performance improvements in some studies, their advantages are not consistent and they do not always outperform more traditional modelling strategies in validation sets or external cohorts [25]. At the same time, studies have shown that deep features extracted from pretrained networks are not necessarily superior to handcrafted radiomic features; in settings with limited sample sizes and high feature dimensionality, the addition of deep features may actually increase redundancy and the risk of overfitting, thereby weakening the model’s generalisation ability [26]. Taken together with the findings of the present study, the integration of clinical variables, conventional PET parameters, and radiomic features may already be sufficient to characterise BMI-related information and achieve a more reasonable balance between predictive performance and model robustness.

SHAP analysis revealed that the model’s key information was derived primarily from first-order statistical features and higher-order texture features, suggesting that the BMI-associated imaging phenotype was not merely characterised by increased bone marrow FDG uptake, but more likely reflected increased grey-level dispersion within the bone marrow, more focal areas of high uptake, and remodelling of spatial heterogeneity. Previous studies have shown that bone marrow involvement in FL on PET/CT may manifest as focal or heterogeneous increased uptake, and that quantitative analysis can improve BMI detection, thereby providing an imaging basis for the contribution of heterogeneity-related features [13]. Orig-FO-IQR represents the degree of dispersion in the grey-level distribution within the ROI; higher values indicate a more dispersed bone marrow metabolic signal, consistent with disruption of the relatively uniform metabolic background of normal bone marrow after tumour infiltration. Wav-HLL-NGTDM-Strength, Wav-HLL-GLRLM-SRHGLE, and LBP3D-m2-GLSZM-SAHGLE collectively indicated more pronounced regional metabolic contrast and enrichment of short-range, small-area high-intensity foci, suggesting that the bone marrow of patients with BMI may contain more scattered or microscopic hypermetabolic infiltrative lesions [15]. LBP3D-m1-GLCM-MCC, in contrast, mainly reflected increased local texture complexity [27]. Because LBP is essentially a grayscale-invariant and rotation-invariant local texture descriptor, its inclusion in the final model suggests that BMI-related information was derived not only from overall uptake intensity but also from alterations in the microscopic texture organisation of bone marrow. From a pathological perspective, this may be related to remodelling of the mixed distribution of normal haematopoietic tissue, adipose components, and tumour components after tumour cell infiltration, thereby appearing as more complex local texture patterns on PET images.

Overall, this study demonstrates that bone marrow involvement in FL is closely associated with higher tumour burden, adverse clinical features, and altered metabolic heterogeneity within the bone marrow; baseline ¹⁸F-FDG PET/CT can provide valuable non-invasive phenotypic information. Compared with the pelvic VOI framework alone, the combined spine-pelvis VOI framework showed superior predictive value after integrating clinical, conventional imaging, and radiomic features, suggesting that broader characterisation of bone marrow phenotypes may improve the identification of BMI.

This study has several limitations. First, as a single-centre retrospective study with a relatively small sample size and no independent external validation, the model’s generalisability requires further evaluation. The initial descriptive screening of clinical and conventional imaging features using Wilcoxon rank-sum and Fisher’s exact tests was performed on the full cohort prior to the training-validation split, which may have introduced minor information leakage; however, all subsequent feature selection and model training steps were conducted exclusively within the training set to mitigate this potential bias. Second, although BMB serves as the clinical reference standard, its single-site sampling nature may underestimate the true extent of bone marrow involvement, thereby affecting the concordance between whole-body bone marrow phenotypes on PET/CT and pathological findings. Third, this study primarily focused on identifying the presence or absence of BMI and did not further explore the extent of bone marrow involvement or its prognostic significance. Finally, although SHAP analysis enhanced the model’s interpretability, the relationship between the relevant imaging features and the biological basis of bone marrow involvement requires further validation.

Conclusions

Bone marrow involvement in follicular lymphoma is associated with higher tumour burden and altered metabolic heterogeneity within the bone marrow, and ¹⁸F-FDG PET/CT provides valuable non-invasive assessment information. Compared with the pelvic VOI framework, the combined spine-pelvis VOI framework showed superior predictive performance for BMI after integrating clinical, conventional imaging, and radiomic features, thereby providing a new methodological basis for the precise identification of bone marrow involvement in follicular lymphoma.

Supplementary Information

Below is the link to the electronic supplementary material.

Authors’ contributions

Li SC and Fan X reviewed the literature and drafted the manuscript. Li SC and He J conceived the idea for the manuscript. Xu YP, He F, Chen QL, and Li AM provided comprehensive perspectives during manuscript preparation. Fan X and He J revised and finalised the manuscript. Fan X and He J played important and indispensable roles in the preparation of the manuscript as co-corresponding authors. All authors read and approved the final manuscript.

Funding

This work was supported by Clinical Trials from the Nanjing Drum Tower Hospital, Affiliated Hospital of Medical School, Nanjing University, No. 2021-LCYJ-MS-11; and Nanjing Drum Tower Hospital National Natural Science Foundation Youth Cultivation Project, No. 2024-JCYJ-QP-15.

Data availability

The datasets used and/or analysed during the current study are available from the corresponding author on reasonable request.

Declarations

Ethics approval and consent to participate

All procedures performed in studies involving human participants were in accordance with the ethical standards of the institutional and/or national research committee and with the 1964 Helsinki Declaration and its later amendments or comparable ethical standards. This retrospective study was conducted at Nanjing Drum Tower Hospital, Affiliated Hospital of Medical School, Nanjing University, approved by the Institutional Ethics Committee (Approval No: 2024-851-01) with waived informed consent.

Consent for publication

This retrospective study used anonymised patient data and was granted a waiver of informed consent by the Institutional Ethics Committee (Approval No: 2024-851-01). All data were de-identified prior to analysis to ensure participant confidentiality.

Competing interests

The authors declare no conflict of interest.