Individualized metabolic brain network alterations in diffuse large B-cell lymphoma: an ¹⁸F-FDG PET study

Abstract

Objective

To investigate cross-sectional differences in individualized metabolic brain network topology between patients with diffuse large B-cell lymphoma (DLBCL) and healthy controls, and to explore the associations between network topology and cognitive performance.

Methods

A total of 104 patients with DLBCL and 33 healthy controls were prospectively enrolled. All participants underwent whole-body ¹⁸F-FDG PET/CT between January 2024 and January 2026. Patients were categorized into baseline and post-chemotherapy groups according to treatment status. Cognitive function was assessed using the Mini-Mental State Examination (MMSE), Montreal Cognitive Assessment (MoCA), and Functional Assessment of Cancer Therapy-Cognitive Function Scale (FACT-Cog). Individual metabolic brain networks were constructed using the Kullback-Leibler Divergence Similarity Estimation (KLSE) method. Graph theory was applied to analyze global and nodal topological properties. Group comparisons and correlation analyses were conducted to evaluate associations between network metrics and cognitive scores (P < 0.05).

Results

We included 137 participants: 104 patients (mean age, 55.6 ± 11.7 years; 57 males) and 33 healthy controls (mean age, 50.3 ± 11.8 years; 17 males). Compared with controls, both baseline and post-chemotherapy patients showed significantly lower MMSE and MoCA scores (P < 0.001). There were no significant differences in cognitive scale scores between the two patient groups. Meanwhile, all groups exhibited small-world properties (γ > 1, λ ≈ 1, σ > 1), with no significant differences in global network metrics (P > 0.05). At the nodal level, baseline patients showed increased nodal degree and efficiency specifically in frontal regions, and decreased betweenness centrality in bilateral parietal regions. In cross-sectional comparisons, post-chemotherapy patients exhibited decreased degree centrality and nodal efficiency in prefrontal regions compared with baseline. In contrast, temporal, occipital, and selected subcortical regions showed increased nodal degree or efficiency. Furthermore, nodal degree, efficiency, and betweenness centrality in these regions were significantly correlated with cognitive scores.

Conclusion

DLBCL patients maintained global network organization but demonstrated specific regional topological changes compared to healthy controls. The observed nodal differences between baseline and post-chemotherapy patients were correlated with cognitive performance. Individualized ¹⁸F-FDG PET/CT assessment of metabolic brain networks offers valuable insight into cognitive impairment in DLBCL.

Introduction

Diffuse large B-cell lymphoma (DLBCL) is one of the most common subtypes of non-Hodgkin lymphoma (NHL), accounting for approximately 30% of all lymphomas [1]. Advances in immunochemotherapy have improved outcomes. Standard regimens such as R-CHOP (rituximab, cyclophosphamide, doxorubicin, vincristine, and prednisone) have achieved cure rates of approximately 60–70% in newly diagnosed patients [2]. While survival outcomes have markedly improved, adverse effects related to disease and treatment have become a clinical concern. Neuropsychological evidence indicates that chemotherapy may induce persistent cognitive sequelae. This is commonly referred to as chemotherapy-related cognitive impairment (CRCI), also known as “chemo-brain”. The reported prevalence of CRCI ranges from 19% to 78%, which may be explained by differences in assessment methods, tumor types, treatment protocols, and follow-up duration. CRCI typically affects several cognitive domains, including learning and memory, information processing speed, and executive function. These impairments may persist for years after treatment, contributing to sustained functional limitations and reduced quality of life [3–7]. The concept of chemo-brain was first proposed in the 1970s [8]. Early studies relied mainly on patients’ subjective cognitive complaints [9, 10]. In the past two decades, systematic neuropsychological and neuroimaging investigations have progressed rapidly [5, 11, 12]. Recent studies have demonstrated that chemotherapy may not only impair memory, attention, and executive function, but may also be associated with multimodal neuroimaging abnormalities involving cerebral metabolism, gray and white matter structure, and functional connectivity [13, 14].

¹⁸F-fluorodeoxyglucose positron emission tomography/computed tomography (¹⁸F-FDG PET/CT) has been widely used for lymphoma diagnosis, staging, assessment of treatment response, and prognostic evaluation [15]. This modality measures the cerebral metabolic rate of glucose (CMRGlc). Thus, ¹⁸F-FDG PET/CT can also characterize brain functional activity, offering a metabolic perspective for investigating cognitive changes beyond tumor assessment [16]. Previous PET studies have shown that chemotherapy can alter global cortical glucose metabolism. Metabolic abnormalities in the frontal lobe, basal ganglia, and cerebellum are closely linked to cognitive decline. These findings suggest that ¹⁸F-FDG PET/CT can objectively reflect functional brain changes related to cognitive impairment in cancer patients [17]. Graph theory models the brain as a complex network, in which brain regions are represented as nodes and interregional associations as edges. This approach allows quantification of topological properties and their relationships with cognitive and behavioral outcomes, providing a robust mathematical method for characterizing brain network organization [18]. Previous studies using magnetic resonance imaging (MRI) have demonstrated the value of network-based analysis for conditions such as mild cognitive impairment, Alzheimer’s disease, and cancer-related cognitive dysfunction [13, 19–23], indicating that network-level metrics may capture cognitive alterations more effectively than isolated regional abnormalities.

However, few studies have systematically compared brain metabolic network characteristics among DLBCL patients at different treatment stages and healthy controls using ¹⁸F-FDG PET/CT. Most have relied on group-level network construction, which can obscure inter-individual variability and limit clinical applicability [24]. Kullback-Leibler divergence similarity estimation (KLSE), which quantifies regional distributional similarity using relative entropy, has been validated for constructing individual-level brain networks [25].

Therefore, the present study aimed to construct individualized metabolic brain networks in baseline and post-chemotherapy DLBCL patients using ¹⁸F-FDG PET/CT combined with the KLSE method. Through the analysis of network topological properties, we sought to investigate cross-sectional differences between DLBCL patients and healthy controls and to explore their associations with cognitive performance. This approach may provide new insights into brain network alterations associated with cognitive impairment in DLBCL patients and help identify potential imaging markers for clinical evaluation.

Methods

Participants

A total of 104 patients with non-central nervous system DLBCL (non-CNS DLBCL) who underwent whole-body ¹⁸F-FDG PET/CT at the Department of Nuclear Medicine, the Second Hospital of Lanzhou University, between January 2024 and January 2026 were prospectively enrolled. Non-CNS involvement was determined through clinical assessment and ¹⁸F-FDG PET/CT examination. Patients with primary central nervous system lymphoma were excluded, and no intracranial lymphoma lesions or other evidence suggestive of CNS involvement were identified in the enrolled patients. Patients were categorized into a baseline group (n = 59), who underwent initial staging before treatment, and a post-chemotherapy group (n = 45), who received response evaluation or restaging within 3 months of completing chemotherapy. This was a cross-sectional study, and the baseline and post-chemotherapy groups consisted of independent participants rather than longitudinal follow-up assessments of the same individuals. The inclusion criteria were as follows: (1) Histopathologically confirmed DLBCL with no evidence of CNS involvement based on clinical assessment and ¹⁸F-FDG PET/CT findings; (2) aged between 18 and 79 years; (3) completion of whole-body ¹⁸F-FDG PET/CT before treatment or within 3 months after completion of standard R-CHOP therapy; (4) availability of cognitive assessment data; (5) complete clinical and imaging data; and (6) right-handedness. The exclusion criteria included: (1) history of other malignancies, traumatic brain injury, neurosurgery, or neuropsychiatric disorders; (2) receipt of intrathecal injection, radiotherapy, or non-standard chemotherapy regimens; and (3) presence of diabetes mellitus.

Additionally, 33 healthy individuals undergoing routine health examinations were recruited as controls and were matched to the patient groups for age, sex, and educational level. Eligible controls had no history of malignant disease, neurological or psychiatric disorders, traumatic brain injury, or brain surgery, and were free of diabetes mellitus and other medical conditions that might affect cerebral metabolism or cognitive function. Only cerebral metabolic PET data were included in the network analysis. This study was approved by the hospital’s Medical Research Ethics Committee (Approval No. 2026A-042), and written informed consent was obtained from all participants or their legal guardians.

Clinical and neuropsychological assessments

Clinical data were systematically collected, including age, sex, education level, Eastern Cooperative Oncology Group (ECOG) performance status, pre-treatment lactate dehydrogenase (LDH) levels, Lugano stage, International Prognostic Index (IPI) score, chemotherapy regimen and cycles, and the interval between the last chemotherapy session and PET/CT examination. At baseline or post-treatment, all patients underwent ¹⁸F-FDG PET evaluation and neuropsychological testing, including the Mini-Mental State Examination (MMSE), the Beijing version of the Montreal Cognitive Assessment (MoCA), and the Functional Assessment of Cancer Therapy-Cognitive Function (FACT-Cog). FACT-Cog consists of four subscales with 37 items: perceived cognitive impairments (PCI), other people’s appraisal (Oth), perceived cognitive ability (PCA), and impact on quality of life (QOL) [26]. The authorized Chinese version was used with permission from the FACIT organization.

PET images acquisition and preprocessing

PET/CT imaging was performed using a Discovery 690 scanner (GE Healthcare, WI, USA). All subjects fasted for 4–6 h before scanning, and fasting blood glucose levels were maintained between 3.9 and 11.0 mmol/L. ¹⁸F-FDG was administered intravenously at a dose of 2.96–5.55 MBq/kg, followed by a 40–60 min rest in a quiet environment. During acquisition, participants were positioned supine, and whole-body and brain scans were obtained sequentially. Brain acquisition time was 5 min per bed position. A low-dose CT scan was first performed with the following parameters: slice thickness of 3.75 mm, tube current of 50–220 mA, tube voltage of 120 kV, and a matrix size of 512 × 512. Subsequently, PET data were acquired in static time-of-flight (TOF) three-dimensional mode with a matrix size of 256 × 256. Images were reconstructed using the VUE Point FX algorithm with a cutoff frequency of 1.0 mm, 18 subsets, and 4 iterations.

Brain PET image preprocessing was performed using MATLAB R2022a (MathWorks, Natick, MA) and SPM12. First, the raw PET data in the Digital Imaging and Communications in Medicine (DICOM) format were converted to the Neuroimaging Informatics Technology Initiative (NIFTI) format. Next, images were manually reoriented and spatially normalized to the Montreal Neurological Institute (MNI) reference space, and finally resampled to a voxel size of 2 mm × 2 mm × 2 mm.

Construction of individual metabolic brain networks

KLSE is a mathematical approach for quantifying the similarity between brain regions and is particularly suitable for complex systems such as the human brain [25]. The brain was first parcellated into 90 regions of interest (ROIs) according to the Anatomical Automatic Labeling (AAL) atlas [27]. The standardized uptake value (SUV) of each ROI was intensity-normalized using the bilateral cerebellar mean SUV as a reference, thereby characterizing regional metabolic attributes. Subsequently, the standardized uptake value ratio (SUVR) of all voxels within each ROI was extracted, and the probability density function (PDF) of SUVR values was estimated using the kernel density estimation (KDE) method proposed by Botev [28]. Based on this, KLSE was calculated to quantify the similarity between the probability density distributions of SUVR values for any pair of ROIs [24, 25], yielding a 90 × 90 individual connectivity matrix for each subject. To avoid isolated nodes and reduce spurious connections, the sparsity threshold was set at values ranging from 0.1 to 0.5 with a step size of 0.01.

After network construction, topological metrics were calculated using GRETNA 2.0 for each subject in the baseline, post-chemotherapy, and control groups across the defined sparsity range. Global metrics included clustering coefficient (Cp), characteristic path length (Lp), normalized clustering coefficient (γ), normalized characteristic path length (λ), small-world parameters (σ), global efficiency (Eg), and local efficiency (Eloc). Nodal metrics included betweenness centrality (Bc), degree centrality (Dc), nodal efficiency (Ne), and nodal local efficiency (NLe). Furthermore, the area under the curve (AUC) across the sparsity range of 10%-50% was calculated for each metric and used in subsequent analyses.

Statistical analysis

Statistical analyses were performed using SPSS 26.0 (IBM, Armonk, NY) and MATLAB R2022a. The normality of continuous variables was assessed using the Shapiro-Wilk test. Normally distributed variables were compared using one-way analysis of variance (ANOVA), whereas non-normally distributed data were analyzed using nonparametric methods. Categorical variables were compared using the chi-square test, and ordinal variables were analyzed using the Kruskal-Wallis H test or the Mann-Whitney U test. Pearson correlation analysis was performed to examine the associations between network parameters and cognitive scores.

For both global and nodal network metrics, one-way ANOVA was conducted across the three groups for each network metric, followed by post hoc pairwise group comparisons using Bonferroni correction to adjust for multiple comparisons between groups. Nodal analyses were performed separately for each brain region and each nodal metric. Given the large number of regional comparisons, nodal-level findings should be interpreted as exploratory. A two-tailed P < 0.05 was considered statistically significant. In addition, network visualization was performed using MATLAB R2022a and BrainNet Viewer, allowing intuitive representation of group differences [29].

Results

Demographic and clinical characteristics

Demographic and clinical characteristics are summarized in Table 1. No significant differences were observed among the three groups in terms of age, sex, or educational level. Compared with the healthy control group, both the baseline and post-chemotherapy groups of patients with DLBCL showed significantly lower MMSE and MoCA total scores (Bonferroni-corrected P < 0.001). In addition, no significant differences were found between the baseline and post-chemotherapy groups in any FACT-Cog subscale scores. These findings suggest that cognitive impairment was present in patients with DLBCL in both the baseline and post-chemotherapy groups.

Table 1.

Demographic and clinical characteristics of all participants

		BS (n = 59)	PC (n = 45)	HC (n = 33)	Statistics	p
Age (years)		56.53 ± 12.28	54.40 ± 10.97	50.30 ± 11.79	F = 2.97	0.055a
Gender (M/F)		30/29	27/18	17/16	χ2 = 0.97	0.616b
Education (years)		10.04 ± 1.97	9.95 ± 1.61	9.75 ± 2.55	H = 5.16	0.076c
Stage(n)	I-II	30	24	/	χ2 = 0.06	0.802b
	III-IV	29	21	/
ECOG(n)	0	33	22	/	χ2 = 0.01	0.905b
	1	19	18	/
	≥ 2	7	5	/
IPI(n)	0–1	33	29	/	z=-0.87	0.387d
	2	10	7	/
	3	7	3	/
	4–5	9	6	/
MMSE score		25.66 ± 3.21	25.07 ± 3.62	28.12 ± 1.22	H = 19.28	<0.001c
MOCA score		20.98 ± 5.35	20.87 ± 4.76	25.61 ± 2.61	H = 24.36	<0.001c
FACT-Cog score	CogPCI	62.54 ± 10.92	63.04 ± 9.17	/	z=-0.11	0.785d
	CogOth	14.32 ± 2.22	14.20 ± 2.57	/	z=-0.44	0.902d
	CogPCA	23.34 ± 6.18	22.40 ± 6.54	/	z=-1.08	0.366d
	CogQOL	12.92 ± 3.51	12.42 ± 3.35	/	z=-0.50	0.416d

BS, baseline group; PC, post-chemotherapy group; HC, healthy control group; MMSE, Mini-Mental State Examination; MoCA, Montreal Cognitive Assessment; FACT-Cog, Functional Assessment of Cancer Therapy-Cognitive Function; CogPCI, perceived cognitive impairments; CogOth, other people’s appraisal; CogPCA, perceived cognitive abilities; CogQOL, impact on quality of life; IPI, International Prognostic Index. p^a, one-way analysis of variance (ANOVA); p^b, chi-square test; p^c, Kruskal-Wallis H test; p^d, Mann-Whitney U test

Comparison of global network topological properties

Across the 10%-50% sparsity range, all three groups showed γ > 1, λ ≈ 1, and σ > 1 (Fig. 1), indicating small-world properties in the metabolic brain networks of all groups. Further comparisons of global network metrics revealed no significant differences among the baseline, post-chemotherapy, and control groups in Cp, Lp, Eg, or Eloc (P > 0.05; Fig. 2).

Fig. 1.

Comparison of small-world properties among groups. Sigma, small-worldness (A); Gamma, normalized clustering coefficient (B); Lambda, normalized characteristic path length (C). BS, baseline group; PC, post-chemotherapy group; HC, healthy control group

Fig. 2.

Comparison of global topological properties among groups. Cp, clustering coefficient of the network (A); Lp, characteristic path length of the network (B); Eg, global efficiency of the network (C); Eloc, local efficiency of the network (D). BS, baseline group; PC, post-chemotherapy group; HC, healthy control group

Comparison of nodal network topological properties

ANOVA followed by post hoc tests revealed significant between-group differences in nodal topological properties of PET metabolic brain networks (Table 2; Fig. 3). Compared with the baseline group, the post-chemotherapy group showed increased Bc in the left inferior frontal gyrus (orbital part) (ORBinf.L) and left postcentral gyrus (PoCG.L), and increased Dc in ORBinf.L and right cuneus (CUN.R). In addition, Ne was increased in CUN.R, and NLe was increased in the right middle frontal gyrus (orbital part) (ORBmid.R). In contrast, decreased Dc and Ne were observed in the bilateral superior frontal gyrus (dorsolateral) (SFGdor.L/R) and the right middle frontal gyrus (MFG.R), along with decreased Ne in the right supplementary motor area (SMA.R), and decreased NLe in the right postcentral gyrus (PoCG.R).

Table 2.

Intergroup differences in nodal network metrics among the three groups

Brain regions	d	P	95%CI
BS vs. HC
Bc
SPG.L	-3.38	0.034	[-0.19, -6.56]
Dc
SFGdor.L	1.77	0.022	[0.19, 3.34]
Ne
SFGdor.L	0.01	0.021	(<0.01,0.02]
NLe
INS.R	0.02	0.013	(<0.01,0.04]
PC vs. BS
Bc
ORBinf.L	6.58	0.029	[0.49, 12.68]
PoCG.L	3.24	0.016	[0.46, 6.03]
Dc
SFGdor.L	-2.07	0.002	[-3.50, -0.63]
SFGdor.R	-1.77	0.010	[-3.21, -0.34]
MFG.R	-1.44	0.018	[-2.70, -0.19]
ORBinf.L	1.25	0.031	[0.08, 2.41]
CUN.R	1.26	0.035	[0.07, 2.46]
SMG.L	2.29	0.036	[0.11, 4.47]
Ne
SFGdor.L	-0.01	0.003	[-0.02, <0.01)
SFGdor.R	-0.01	0.019	[-0.02, <0.01)
MFG.R	-0.01	0.025	[-0.02, <0.01)
SMA.R	-0.01	0.018	[-0.01, <0.01)
CUN.R	0.01	0.027	(<0.01,0.02]
NLe
ORBmid.R	0.03	0.035	(<0.01,0.05]
PoCG.R	-0.01	0.014	[-0.01, <0.01)
PC vs. HC
Dc
CAL.R	2.31	0.044	[0.046, 4.57]
ANG.L	-2.20	0.016	[-4.08, -0.31]
PCUN.R	-1.87	0.004	[-3.26, -0.48]
TPOmid.R	0.38	0.025	[0.04, 0.72]
Ne
CAL.R	0.02	0.025	(<0.01,0.03]
ANG.L	-0.01	0.031	[-0.02, <0.01)
PCUN.R	-0.01	0.006	[-0.02, <0.01)
NLe
INS.R	0.02	0.046	(<0.01,0.04]
PUT.R	0.02	0.027	(<0.01,0.04]

BS, baseline group; PC, post-chemotherapy group; HC, healthy control group; Bc, betweenness centrality; Dc, degree centrality; Ne, nodal efficiency; NLe, nodal local efficiency; brain regions are defined according to the AAL atlas. P, Bonferroni-corrected p values; d, between-group difference (positive values indicate higher values in the former group); 95% CI, 95% confidence interval

Fig. 3.

Comparison of nodal topological properties between groups. (A–D) Baseline group vs. post-chemotherapy group; (E–H) baseline group vs. healthy control group; (I–K) post-chemotherapy group vs. healthy control group. Red nodes represent regions with significantly higher nodal values, whereas blue nodes represent regions with significantly lower nodal values

Relative to the healthy control group, the baseline group exhibited increased Dc and Ne in the left superior frontal gyrus (dorsolateral) (SFGdor.L) and increased NLe in the right insula (INS.R), whereas decreased Bc was observed in the left superior parietal gyrus (SPG.L). The post-chemotherapy group showed increased Dc in the right temporal pole (middle temporal gyrus) (TPOmid.R), as well as increased Dc and Ne in the right calcarine fissure and surrounding cortex (CAL.R). Increased NLe was observed in INS.R and right lenticular nucleus (putamen) (PUT.R), while decreased Dc and Ne were found in the left angular gyrus (ANG.L) and right precuneus (PCUN.R).

Associations between nodal topological properties and cognitive scores

As shown in Fig. 4, significant correlations between nodal topological properties and cognitive performance were observed. In the baseline group, the Bc of SPG.L was positively correlated with MMSE scores (r = 0.294, P = 0.028). In contrast, the Dc and Ne of MFG.R were negatively correlated with MMSE scores (r = -0.286, P = 0.033; r = -0.307, P = 0.022). In addition, the Bc of ORBinf.L, as well as the Dc and Ne of SFGdor.L, showed positive correlations with PCA scores (r = 0.387, P = 0.003; r = 0.313, P = 0.019; r = 0.327, P = 0.014). The Dc of ORBinf.L was also positively correlated with QOL scores (r = 0.290, P = 0.030), while the NLe of PoCG.R was positively correlated with PCI scores (r = 0.267, P = 0.047).

Fig. 4.

Correlation analyses in the baseline (A) and post-chemotherapy (B) groups. MMSE, Mini-Mental State Examination; PCA, perceived cognitive abilities; QOL, quality of life; PCI, perceived cognitive impairment; Bc, betweenness centrality; Dc, degree centrality; Ne, nodal efficiency; NLe, nodal local efficiency

In the post-chemotherapy group, nodal topological properties were primarily characterized by negative associations with QOL scores. The Dc and Ne of SFGdor.L/R were negatively correlated with QOL scores (Dc: r = -0.515, P < 0.001; r = -0.444, P = 0.003; Ne: r = -0.547, P < 0.001; r = -0.465, P = 0.002). In addition, the Ne of SMA.R was negatively correlated with QOL scores (r = -0.346, P = 0.025).

Discussion

In the present study, a KLSE-based approach was applied for the first time to construct individualized PET metabolic brain networks in baseline and post-chemotherapy DLBCL patients. Graph theory analysis was further employed to systematically evaluate alterations in network topological properties and their associations with cognitive function. The findings indicated that, although DLBCL patients maintained the typical small-world property, regional alterations in nodal topological properties were observed in the frontal, parietal, occipital, and limbic systems. Importantly, these alterations were closely associated with both objective cognitive performance and subjective cognitive complaints.

Unlike previous studies based on functional magnetic resonance imaging (fMRI) or group-level structural covariance networks [25, 30–32], the present study applied the KLSE method to PET metabolic data in lymphoma patients for the first time. Importantly, this approach enabled characterization of brain network alterations at the individual level and provided a more refined perspective on mechanisms underlying cognitive impairment in DLBCL patients. In this study, MMSE and MoCA scores were significantly reduced in both baseline and post-chemotherapy DLBCL patients compared with healthy controls (P < 0.001), suggesting the presence of cognitive impairment in patients with DLBCL. Importantly, cognitive impairment was observed in both the baseline and post-chemotherapy groups, suggesting that mechanisms beyond treatment-related factors may contribute to cognitive dysfunction in DLBCL patients. Previous studies have indicated that hematological malignancies themselves may affect central nervous system function through systemic inflammation, cytokine dysregulation, oxidative stress, metabolic abnormalities, and blood-brain barrier disruption. Pro-inflammatory cytokines and microglial activation may further interfere with neuronal metabolism and functional connectivity, thereby contributing to cognitive impairment [33–35]. In addition, disease burden, fatigue, anxiety, and psychological distress, which are frequently observed in lymphoma patients, may also influence cognitive performance [36, 37]. Previous studies have suggested that chemotherapy may be associated with alterations in key brain networks, such as the prefrontal-cerebellar system (PCS) [33], while the tumor itself may also affect central nervous system function through mechanisms including systemic inflammation. Taken together, the observed cognitive impairment in DLBCL patients may be associated with multiple disease- and treatment-related factors.

From a global network perspective, small-world topology provides the structural basis for efficient information integration and segregation in the brain. In the present study, all three groups maintained small-world properties, and no significant differences were observed in global topological metrics, consistent with previous reports [31, 38]. This finding suggests that large-scale network organization remains relatively preserved despite disease burden and treatment exposure, possibly reflecting compensatory or adaptive mechanisms at the macroscopic level. Notably, reduced cognitive performance was observed despite preserved global topology, suggesting a potential dissociation between network architecture and functional outcomes. Therefore, the observed network differences in the present study were primarily reflected at the nodal level rather than in global network organization. Several explanations may account for this observation. First, global network disruption is more likely to emerge at later stages of impairment, whereas early or moderate dysfunction may not affect overall topology [32, 39]. Second, nodal topological abnormalities may be more sensitive to disease- and treatment-associated disturbances in the brain network than global topological metrics [27]. Accordingly, nodal metrics may be more sensitive indicators of subtle network abnormalities than global parameters.

Nodal topological measures reflect the functional roles of specific brain regions and their efficiency in information processing. In the present study, nodal topological abnormalities were observed in DLBCL patients before chemotherapy. Compared with controls, reduced Bc in SPG.L may indicate reduced interregional information transfer efficiency, which has been implicated in attention and working memory processes. In contrast, increased Dc and Ne in SFGdor.L, along with elevated NLe in INS.R, may represent enhanced local processing and integration, potentially reflecting adaptive responses to disease-related burden [40–42]. Previous research has shown that hematological malignancies may be associated with neuronal metabolic abnormalities through mechanisms involving systemic inflammation, blood-brain barrier dysfunction, and microglial activation, which is consistent with the current findings [1].

Cross-sectional comparisons revealed more extensive regional differences in nodal topological properties in the post-chemotherapy cohort than in the baseline cohort. Reduced Dc and Ne in SFGdor.L/R and MFG.R may reflect relatively reduced connectivity strength and information transmission efficiency in prefrontal regions implicated in executive function. This observation may be consistent with the decline in cognitive performance observed in DLBCL patients. Previous studies have suggested that the prefrontal cortex may be vulnerable to treatment-related neurobiological influences, including altered synaptic plasticity and disrupted functional connectivity [34, 43–47]. Meanwhile, increased Dc or Ne in ORBinf.L and CUN.R may reflect recruitment of additional regions during cognitive processing when core cognitive networks exhibit reduced efficiency [23, 33, 38]. This pattern suggests that functional redistribution across brain regions may occur, which may help explain the coexistence of preserved global topology and cognitive impairment.

In addition, differences in nodal topological properties between the baseline and post-chemotherapy groups were region-specific. Reduced Dc and Ne in ANG.L and PCUN.R suggest lower integration efficiency within the default mode network. Conversely, increased NLe in INS.R, PUT.R, and CAL.R reflect enhanced local processing efficiency, possibly related to functional reorganization of sensorimotor and salience networks. When information processing efficiency in core networks is reduced, these regions may participate more extensively in information integration. Such findings may reflect adaptive or compensatory neural responses, although their functional significance remains uncertain [48–51].

Correlation analysis further highlighted distinct association patterns between groups. In the baseline group, frontal nodal properties were positively associated with FACT-Cog scores, suggesting potential compensatory mechanisms. In contrast, negative correlations involving MFG.R and MMSE scores may reflect inefficient information processing or limited compensatory capacity [31, 32, 38]. In the post-chemotherapy group, nodal properties in SFGdor.L/R and SMA.R were negatively correlated with QOL scores, which may indicate altered neural resource utilization. Increased recruitment of neural resources required to maintain cognitive function may be associated with greater cognitive burden and reduced perceived quality of life [52, 53]. Similar associations between network efficiency and cognitive performance have also been reported in other cancer populations [38, 54].

This study has several limitations. First, it was designed as a cross-sectional analysis, and longitudinal follow-up within the same individuals was not performed. Second, the sample size was relatively small, and a fully matched design was not achieved, which may have limited statistical power and restricted subgroup analyses. Third, brain metabolic data were derived from routine whole-body PET/CT examinations performed for oncological purposes rather than dedicated neuroimaging protocols, which may introduce variability in voxel-wise and network-based analyses due to differences in acquisition conditions and image resolution. In clinical practice, patients with lymphoma do not routinely undergo dedicated brain PET imaging during follow-up. Therefore, brain data were acquired according to standard whole-body PET/CT protocols, with consideration of clinical feasibility and reduced radiation exposure. Fourth, although Bonferroni correction was applied for post hoc group comparisons, no additional correction for multiple testing was performed across all nodal and correlation analyses. Given the large number of brain regions, nodal metrics, and cognitive measures examined, the possibility of false-positive findings cannot be completely excluded. Therefore, the regional nodal alterations and their associations with cognitive performance should be regarded as exploratory and interpreted with appropriate caution until confirmed in larger independent cohorts. Finally, although strict inclusion and exclusion criteria were applied, the potential influence of the disease itself, as well as factors such as anxiety, depression, and fatigue, cannot be entirely excluded. Therefore, the present findings should be interpreted as cross-sectional associations rather than as direct evidence of chemotherapy-induced longitudinal alterations in brain networks.

Conclusions

Despite these limitations, the present study constructed individualized PET metabolic brain networks in DLBCL patients using the KLSE method and characterized their topological features, demonstrating preserved global organization alongside regional nodal differences. Furthermore, altered nodal topological properties were associated with cognitive impairment. Cross-sectional comparisons revealed more extensive regional nodal differences, particularly in prefrontal and other brain regions, in the post-chemotherapy cohort than in the baseline cohort. These findings provide additional neuroimaging evidence of cognitive impairment in DLBCL patients and suggest that individualized brain metabolic network analysis may be a promising tool for cognitive evaluation. Future studies with larger sample cohorts and longitudinal designs are warranted to further validate and clarify these findings.

Abbreviations

DLBCL

Diffuse large B-cell lymphoma

CRCI

Chemotherapy-related cognitive impairment

¹⁸F-FDG PET/CT

¹⁸F-fluorodeoxyglucose Positron emission tomography/computed tomography

KLSE

Kullback–Leibler divergence similarity estimation

MRI

Magnetic resonance imaging

CMRGlc

Cerebral metabolic rate of glucose

MMSE

Mini-Mental State Examination

MoCA

Montreal Cognitive Assessment

FACT-Cog

Functional Assessment of Cancer Therapy–Cognitive Function

PCI

Perceived cognitive impairments

PCA

Perceived cognitive abilities

QOL

Quality of life

Oth

Other people’s appraisal

SUV

Standardized uptake value

SUVR

Standardized uptake value ratio

KDE

Kernel density estimation

PDF

Probability density function

AAL

Automated Anatomical Labeling

ROI

Region of interest

Cp

Clustering coefficient

Lp

Characteristic path length

Eg

Global efficiency

Eloc

Local efficiency

Gamma

Normalized clustering coefficient

Lambda

Normalized characteristic path length

Sigma

Small-worldness

Bc

Betweenness centrality

Dc

Degree centrality

Ne

Nodal efficiency

NLe

Nodal local efficiency

AUC

Area under the curve

SPG

Superior parietal gyrus

SFGdor

Superior frontal gyrus, dorsolateral

MFG

Middle frontal gyrus

ORBinf

Inferior frontal gyrus, orbital part

INS

Insula

CUN

Cuneus

SMA

Supplementary motor area

Author contributions

Y.G. and F.Y. designed this study and collected, processed, and analyzed the data; H.Z. collected the data; J.X. analyzed the data; and J.L. designed the study and contributed to editing and reviewing the manuscript. The authors read and approved the final manuscript.

Funding

The financial support for this work was provided by the Science and Technology Program of Gansu Province (26YFFA030). The funding organizations were not involved in the study’s design, data collection, analysis, interpretation, or the preparation of the manuscript.

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

This study was conducted in accordance with the Declaration of Helsinki. The study protocol was approved by the Medical Research Ethics Committee of the hospital (Approval No. 2026A-042), and written informed consent was obtained from all participants or their legal guardians.

Consent for publication

Not applicable.

Competing interests

The authors declare no competing interests.