Multiscale Brain Structural Network Reorganization in Diabetic Retinopathy Using Regional Radiomics Similarity Networks Authors: Die Hu, Rui-Yang Hu, Yu-Yue Cheng, Fang-Liang Yu, Xin Huang Published: 2026 ABSTRACT AbstractBackgroundDiabetic retinopathy (DR), a common microvascular complication of diabetes, is increasingly viewed as a systemic disorder involving the retina–brain axis. Although neuroimaging studies have reported brain abnormalities in DR, large-scale network reorganization and its biological substrates remain unclear.MethodsThis cross-sectional study included 47 patients with type 2 diabetes mellitus and non-proliferative diabetic retinopathy (T2DM-NPDR) and 45 matched healthy controls. Structural MRI data were used to construct individualized regional radiomics similarity networks (R2SNs). Group differences in network architecture were examined using multiscale network analysis and gradient mapping. Epicenter mapping identified candidate regions related to network abnormalities. Imaging–transcriptomic analysis based on the Allen Human Brain Atlas, cell-type enrichment analysis, and spatial associations with normative neurotransmitter receptor and transporter maps were performed. Machine learning classifiers and an interpretable graph neural network assessed the discriminative value of R2SN-derived features.ResultsPatients with T2DM-NPDR showed widespread cortical and subcortical alterations in R2SN organization. Gradient analysis revealed a systematic shift in macroscale hierarchy, indicating disrupted cortical hierarchical organization. Epicenter mapping highlighted frontal regions, especially the inferior frontal gyrus, middle frontal gyrus, and orbitofrontal cortex, with additional involvement of the insula and parahippocampal gyrus, suggesting abnormalities beyond primary visual pathways and implicating higher-order integrative systems. Imaging–transcriptomic analysis showed that the spatial pattern of R2SN alterations was associated with gene expression profiles enriched in neuronal and glial cell populations. R2SN abnormalities also showed significant spatial correspondence with normative neurotransmitter receptor and transporter maps. Predictive models demonstrated that R2SN-derived features could distinguish patients with T2DM-NPDR from healthy controls.ConclusionsT2DM-NPDR is associated with multiscale structural brain network alterations beyond the visual system. Integrating radiomics-based network modeling, gradient analysis, molecular mapping, and graph-based prediction provides a framework for characterizing brain network disruption and potential biological associations in T2DM-NPDR.Keywords: diabetic retinopathy, regional radiomics similarity network, brain network, magnetic resonance imaging, radiomics, central nervous system, brain structural changes Introduction Diabetic retinopathy (DR) is one of the most common microvascular complications of diabetes and a leading cause of visual impairment and blindness among working-age adults worldwide.1,2 In 2020, more than 103 million individuals were affected by DR globally, and this number is projected to increase to approximately 161 million by 2045.3 Despite improvements in clinical management and gradual reductions in diabetes-related vision loss, the increasing prevalence of diabetes continues to impose a substantial public health burden.4 Beyond its ocular manifestations, growing evidence suggests that DR reflects a broader systemic microvascular disorder that may also involve the central nervous system.5 The retina and the brain share common embryological origins and exhibit similar neurovascular architectures.6,7 In particular, the blood–retinal barrier closely resembles the blood–brain barrier in both structure and function.8 Consequently, microvascular abnormalities observed in the retina may mirror analogous vascular and neurovascular alterations in the brain.9 Chronic hyperglycemia and insulin resistance, key pathological features of diabetes, can induce endothelial dysfunction and widespread microvascular injury, potentially contributing not only to retinal pathology but also to cerebral microcirculatory disturbances and brain structural alterations.10,11 Consistent with this view, previous studies have reported changes along the visual pathway extending from the retina to the visual cortex, suggesting that DR may affect distributed neural systems beyond the eye. Together, these findings highlight the concept of a retina–brain axis and underscore the need for whole-brain approaches to characterize DR-related neural alterations.11,12 Importantly, diabetes mellitus is a systemic metabolic disease that affects both microvascular and macrovascular systems. Although DR is primarily characterized by retinal microvascular lesions, diabetes-related brain alterations may also involve cerebral microvascular dysfunction, macrovascular pathology, atherosclerosis, ischemia, and reduced cerebral blood flow.10,12,13 This broader vascular context provides an important background for interpreting brain changes observed in patients with DR. Accumulating neuroimaging studies have provided increasing evidence of widespread brain alterations associated with DR, including changes in gray matter structure, white matter microstructure, functional activity, cerebral perfusion, and metabolic processes.14–16 Using diverse MRI modalities such as voxel-based morphometry, diffusion tensor imaging, resting-state functional MRI, arterial spin labeling, and magnetic resonance spectroscopy, previous investigations have consistently identified abnormalities not only within visual pathways but also in brain regions supporting cognitive, emotional, and sensorimotor functions.17–19 However, most existing studies have focused on regional alterations or modality-specific findings, offering limited insight into how distributed brain regions are structurally coordinated as an integrated network. Given that the brain operates as a complex interconnected system, characterizing DR-related alterations from a network-level perspective may provide a more comprehensive understanding of disease-related brain reorganization. Structural brain network approaches, including structural covariance networks and morphometric similarity networks (MSNs), have contributed valuable insights into coordinated brain organization in both healthy populations and neurological disorders.20–22 While structural covariance networks are typically constructed at the group level based on inter-subject variability, MSN approaches can be implemented at the individual level, providing a more personalized representation of brain organization.23,24 Nevertheless, these approaches primarily rely on macroscopic morphometric features such as cortical thickness or gray matter volume and may therefore overlook subtle microstructural variations within brain regions.25–27 To address these limitations, regional radiomics similarity networks (R2SNs) have recently been proposed as an individual-level framework for modeling macroscale structural brain networks.28 By leveraging high-dimensional radiomic features derived from structural MRI, R2SN captures the distributional and textural characteristics of voxel-wise signal intensity within each brain region.29–31 Interregional similarity in these microstructure-sensitive features enables the construction of individualized structural brain networks that may be particularly sensitive to subtle and spatially heterogeneous brain alterations.27,32 Beyond traditional graph-theoretical metrics, recent advances in network neuroscience have emphasized that large-scale brain organization is also structured along continuous hierarchical gradients.33–35 These gradients represent smooth transitions in connectivity and microstructure across the cortex and can be embedded within a low-dimensional manifold space.33,36,37 Such gradient representations provide a powerful framework for understanding macroscale cortical organization and have been widely applied to characterize hierarchical brain architecture in health and disease. Applying gradient analysis to R2SN may therefore provide additional insights into whether DR is associated with alterations in macroscale structural hierarchy beyond localized network abnormalities. To further examine the network-level organization of disease-related alterations, epicenter analysis offers a complementary framework for identifying candidate regions that may anchor large-scale brain network disruptions.38,39 The central premise of this approach is that if a brain region acts as a network epicenter, its normative connectivity profile should resemble the spatial distribution of disease-related abnormalities across the brain.40,41 By integrating regional abnormality maps with normative connectivity architecture, epicenter mapping can reveal potential anchor regions that organize network-level alterations and provide insight into the spatial architecture of disease-related brain changes.38 Understanding the biological mechanisms underlying imaging-derived brain alterations remains an important challenge.42 The Allen Human Brain Atlas (AHBA) provides a comprehensive transcriptomic resource containing whole-brain gene expression profiles derived from thousands of anatomically distributed tissue samples and covering more than 20,000 genes.43 Integrating AHBA-derived spatial gene expression data with neuroimaging findings enables imaging–transcriptomic analyses that link molecular processes with large-scale brain organization.44 Importantly, increasing evidence suggests that genetic factors contribute to the development and progression of DR, with multiple genes and biological pathways implicated in its pathophysiology.45–47 Imaging–transcriptomic approaches may therefore provide valuable insight into the molecular substrates underlying DR-related brain alterations. To further refine biological interpretation, cell-type–specific expression analyses can be applied to determine whether disease-related gene signatures are preferentially associated with particular neural cell populations.48,49 In addition to transcriptional regulation, the organization of brain networks is strongly influenced by neurochemical systems. Neurotransmitter receptors and transporters play key roles in regulating synaptic signaling, neuronal excitability, and circuit-level plasticity, thereby shaping the stability and vulnerability of large-scale brain networks.50,51 Normative maps of neurotransmitter receptor and transporter densities derived from positron emission tomography (PET) and single-photon emission computed tomography (SPECT) provide an opportunity to examine whether disease-related imaging alterations are constrained by intrinsic neurochemical architecture.51,52 Examining the spatial correspondence between these molecular maps and DR-related network alterations may therefore provide complementary insights into the neurochemical substrates underlying structural brain reorganization. Finally, machine learning and graph neural network approaches provide complementary tools for evaluating whether imaging-derived network features contain discriminative information at the individual-subject level. In particular, GNNs are well suited for modeling graph-structured brain network data and may help identify disease-relevant network patterns.53–55 In this study, these predictive models were used to assess the potential classification value of R2SN-derived features in distinguishing patients with T2DM-NPDR from healthy controls. In the present study, we aimed to characterize large-scale brain structural network alterations in patients with type 2 diabetes mellitus and non-proliferative diabetic retinopathy (T2DM-NPDR) compared with healthy controls using a multiscale network neuroscience framework based on R2SNs. Specifically, we examined regional R2SN alterations, macroscale hierarchical organization, candidate network epicenters, potential transcriptomic and neurochemical associations, and the discriminative value of R2SN-derived features. This multiscale strategy was designed to provide a comprehensive characterization of structural brain network alterations in patients with T2DM-NPDR. Methods and Materials Participants This study enrolled 92 participants, including 47 patients with type 2 diabetes mellitus and non-proliferative diabetic retinopathy (T2DM-NPDR) and 45 healthy controls matched for age, sex, and handedness. Inclusion Criteria for patients with T2DM-NPDR: (1) Fasting plasma glucose ≥ 7.0 mmol/L, random plasma glucose ≥ 11.1 mmol/L, or 2-hour glucose level ≥ 11.1 mmol/L. (2) Non-proliferative diabetic retinopathy (NPDR) with features such as microaneurysms, hard exudates, and retinal hemorrhages. (3) All patients with T2DM-NPDR underwent fundus photography and fluorescein angiography. Exclusion Criteria for patients with T2DM-NPDR: (1) Presence of proliferative diabetic retinopathy (PDR) with complications like retinal detachment or vitreous hemorrhage. (2) Coexisting ocular conditions such as cataracts, glaucoma, high myopia, or optic neuritis. (3) Concomitant diabetes-related systemic microvascular or neurological complications, including diabetic nephropathy, defined as a urinary albumin-to-creatinine ratio >30 mg/g persisting for more than three months, or diabetic neuropathy, defined by symptoms such as pain, numbness, or paresthesia together with abnormal nerve conduction findings. (4) Severe neurological disorders, including Alzheimer’s disease (AD), Parkinson’s disease (PD), frontotemporal dementia (FTD), vascular dementia (VaD), or other neurodegenerative diseases. (5) Use of medications that may significantly alter brain structure or function (eg, neuroleptics, antidepressants, or antiepileptics). The inclusion and exclusion criteria for healthy controls were: (1) Never diagnosed with diabetes, with fasting plasma glucose between 3.9 mmol/L and 5.6 mmol/L, random plasma glucose < 11.1 mmol/L, and HbA1c levels < 6.5%. (2) No history of ocular diseases (eg, myopia, cataracts, glaucoma, optic neuritis, or retinal degeneration). (3) Best-corrected visual acuity ≥ 1.0 (decimal) in both eyes. (4) No history of ocular surgery or mental disorders. (5) Severe neurological disorders (as mentioned for patients with T2DM-NPDR) that could affect brain structure were excluded. Overview of the Analytical Framework The overall analytical workflow of this study is illustrated in Figure 1. Individualized R2SNs were first constructed from structural MRI data to capture microstructure-sensitive patterns of interregional similarity across the brain. Regional alterations in R2SN organization were then examined, followed by gradient analysis to characterize the hierarchical architecture of structural brain networks. Epicenter mapping was performed by integrating regional abnormality patterns with normative connectivity architecture to identify candidate regions associated with network-level abnormalities. To explore potential biological associations of the imaging findings, imaging–transcriptomic analyses based on the Allen Human Brain Atlas were conducted, followed by cell-type enrichment analyses and spatial associations with normative neurotransmitter receptor and transporter density maps derived from PET and SPECT imaging. Finally, to evaluate the discriminative capability of R2SN-derived features, machine learning classifiers and an interpretable graph neural network framework were applied for disease prediction. This workflow was designed to characterize brain structural network alterations from imaging-derived network abnormalities to potential biological associations and predictive modeling. Overview of the analytical workflow. (A) Individualized R2SNs were constructed from structural MRI data, and regional case–control T-value maps were generated to characterize R2SN alterations in T2DM-NPDR. (B) Gradient analysis assessed macroscale hierarchical organization. (C) Epicenter mapping identified candidate anchor regions associated with distributed R2SN abnormalities. (D) PLS analysis linked R2SN alteration maps with AHBA-derived gene expression profiles. (E) Gene enrichment and cell-type-specific analyses characterized potential transcriptomic and cellular associations. (F) Neurotransmitter receptor/transporter mapping examined spatial associations with normative neurochemical systems. (G) Machine learning with SHAP interpretation evaluated R2SN-derived regional features. (H) GNN analysis evaluated graph-structured R2SN representations for classification. Data Acquisition Structural magnetic resonance imaging (MRI) data were acquired using a 3.0-Tesla scanner (Discovery MR750W; GE Healthcare, Milwaukee, WI, USA) with an eight-channel phased-array head coil. High-resolution T1-weighted images were obtained through a three-dimensional spoiled gradient-recalled echo (3D-SPGR) sequence in the sagittal orientation. The structural acquisition consisted of approximately 176 contiguous slices covering the entire brain. Imaging parameters were set as follows: repetition time (TR) = 8.5 ms, echo time (TE) = 3.3 ms, slice thickness = 1.0 mm with no interslice gap, acquisition matrix = 256 × 256, field of view (FOV) = 240 × 240 mm2, and flip angle = 12°. This imaging protocol provided isotropic, whole-brain anatomical images with high spatial resolution, which were suitable for subsequent preprocessing and structural analyses. All images were visually inspected by experienced researchers to ensure the absence of major motion artifacts or structural abnormalities prior to further analysis. Data Processing Prior to network construction, all T1-weighted images underwent preprocessing. For each participant, the anatomical image was nonlinearly registered to the Montreal Neurological Institute (MNI) standard template using Advanced Normalization Tools (ANTs; https://github.com/ANTsX/ANTs). This spatial normalization step ensured anatomical comparability across subjects. After normalization, the brain was segmented into predefined regions according to the Brainnetome Atlas, which provides a connectivity-informed parcellation scheme consisting of 246 regions of interest (ROIs), including 210 cortical and 36 subcortical regions.56 The use of a standardized template and atlas-based parcellation ensured anatomical correspondence across subjects and facilitated reproducible region-wise analyses. Radiomic Feature Extraction Radiomic descriptors were extracted from the normalized T1-weighted images within each Brainnetome atlas ROI. Initially, a total of 47 quantitative features were calculated for every region. These descriptors included first-order intensity statistics as well as higher-order texture features that capture spatial relationships among gray-level values, thereby reflecting both regional intensity distributions and structural heterogeneity within each ROI. To minimize redundancy and enhance the stability of the feature set, highly correlated features were excluded following the procedure described in the original R2SN framework. Specifically, features with Pearson correlation coefficients greater than 0.9 were removed. After this filtering step, 25 radiomic features remained and were retained for subsequent network construction.28,57 This feature selection strategy is consistent with prior R2SN studies and has been shown to improve feature robustness and reduce redundancy in high-dimensional radiomic representations. Although radiomic features may be sensitive to regional size and signal variability, the use of atlas-based parcellation and within-subject normalization mitigates potential confounding effects related to interregional heterogeneity. Individualized R2SN Construction and Group Comparisons To facilitate comparability of radiomic measurements across brain regions within each individual, the retained features were first rescaled using min–max normalization across all ROIs on a subject-wise basis. Based on these normalized feature vectors, individualized R2SNs were generated by calculating Pearson correlation coefficients between every pair of ROIs. This procedure yielded a 246 × 246 similarity matrix for each participant, where nodes correspond to Brainnetome atlas regions and edges represent the similarity between regional radiomic profiles. The resulting subject-specific R2SN matrices served as the basis for subsequent analyses. Regional network features were extracted from each matrix, and group differences between patients with T2DM-NPDR and healthy controls were evaluated using region-wise two-sample t-tests. To account for multiple comparisons across the 246 Brainnetome regions, p values were corrected using the Benjamini–Hochberg false discovery rate (FDR), with q < 0.05 considered statistically significant. The resulting t-statistics were used to map the spatial pattern of R2SN alterations across the brain. Brain regions showing significant between-group differences were used only for descriptive projection onto the Yeo 7 functional systems. Subsequent gradient, imaging–transcriptomic, and neurochemical analyses were performed on whole-brain maps across all 246 Brainnetome regions without prior ROI selection. Systems-Level Characterization of R2SN Abnormalities To obtain an intuitive overview of large-scale brain network patterns associated with T2DM-NPDR-related R2SN alterations, system-level characterization and visualization were conducted using the Yeo 7-network functional framework together with a word cloud representation derived from group-level R2SN analyses.58 First, group-mean R2SN matrices were calculated for the T2DM-NPDR and healthy control groups by averaging the individualized R2SN matrices within each group. Subsequently, edge-wise statistical comparisons between groups were performed to generate a case–control R2SN t-statistic map. Based on this map, brain regions and connections showing significant alterations were identified and visualized in three-dimensional brain space. To further examine how these R2SN alterations were distributed across large-scale functional systems, abnormal R2SN measures and regional residual values were summarized according to the Yeo 7-network functional parcellation. In addition, affected regions were projected onto the Brain Reward System to explore potential involvement of reward-related circuits. Functional network annotations and system-level descriptors linked to abnormal R2SN patterns were then compiled. A word cloud visualization was subsequently created, in which the size of each term reflected the relative frequency or prominence of the corresponding functional system. This representation was intended as an exploratory and descriptive summary of the dominant functional systems implicated in R2SN abnormalities rather than as a formal statistical inference. Gradient Analysis of R2SN To examine the macroscale topological organization of the R2SN connectome, gradient analysis was performed using the diffusion map-embedding technique. Diffusion map-embedding is a nonlinear dimensionality reduction approach that models a diffusion process on the data to capture the intrinsic geometric structure of high-dimensional connectivity patterns. By transforming pairwise distances between regions into transition probabilities, this method identifies low-dimensional gradients that reflect the principal axes of variation in network organization. Gradient computation was performed using the BrainSpace toolbox (https://brainspace.readthedocs.io/en/latest/).36 Eigenvalue decomposition of the diffusion operator was applied to obtain the principal gradients describing the large-scale organization of the R2SN connectome. In the present study, gradient estimation was performed directly on the individualized R2SN matrices from T2DM-NPDR patients and healthy controls, without using an external reference template. The resulting gradients provided a low-dimensional representation of the structural similarity network, allowing group comparisons of macroscale topographical organization between T2DM-NPDR patients and healthy controls. Epicenter Analysis of R2SN Network Alterations Epicenter analysis was performed to identify brain regions whose intrinsic structural connectivity patterns most closely matched the spatial distribution of T2DM-NPDR-related R2SN alterations. Normative connectivity profiles were derived from the group-averaged R2SN of healthy controls (see R2SN construction above). For each ROI, the normative connectivity profile was defined as a vector representing its structural similarity with all other brain regions. To quantify epicenter effects at the individual-subject level, a subject-specific abnormality map was defined for each participant based on the regional R2SN alteration pattern. For each participant and for each ROI, a Spearman correlation was then computed between the ROI’s normative connectivity profile and the participant-specific abnormality map, thereby quantifying the degree to which that region’s normative connectivity pattern resembled the individual spatial distribution of R2SN abnormalities. Spearman correlation was adopted to provide a robust measure of association that is less sensitive to potential non-normality and outliers in the data. The resulting correlation coefficients were subsequently converted using Fisher’s r-to-z transformation to yield subject-level regional goodness-of-fit (GOF) values. This procedure generated one GOF value per ROI for each participant. At the group level, one-sample t-tests were performed on the GOF values of each ROI across participants to identify regions showing consistently significant epicenter effects. To account for multiple comparisons across regions, false discovery rate (FDR) correction was applied, with q < 0.05 considered statistically significant. Regions exhibiting significant GOF effects were considered candidate structural-network epicenters, and the region with the highest group-level statistic was defined as the top-ranked epicenter. To further account for spatial autocorrelation inherent in brain maps, spatial permutation (spin) tests (n = 1000) were performed by randomly rotating the cortical abnormality maps on the spherical surface while preserving spatial contiguity. Empirical correlation values were then compared against the corresponding null distributions to derive spin-based significance estimates. To further characterize the large-scale functional context of the identified epicenter, its normative whole-brain functional connectivity (FC) profile was examined using a standard functional connectivity template derived from healthy individuals. FC strengths were summarized within the Yeo functional network parcellation by averaging connectivity values across regions belonging to each network. Networks were subsequently ranked according to their mean FC with the epicenter. This functional projection analysis was conducted for interpretative purposes only and did not contribute to the statistical identification of epicenter regions. It should be noted that epicenter mapping identifies regions whose connectivity profiles resemble the spatial pattern of abnormalities and does not imply causal or temporal primacy. Imaging-Transcriptomic and Neurochemical Analyses Regional gene expression data were obtained from the AHBA, comprising 3702 postmortem samples from six neurologically normal adult donors. Gene expression profiles were processed and mapped to Brainnetome atlas regions using the abagen toolbox, yielding regional expression measures for 15,633 genes corresponding to the R2SN t-statistic map. Standard abagen preprocessing procedures were applied, including probe reannotation, filtering of low-quality samples, probe selection, normalization, assignment of samples to atlas regions, and aggregation of regional expression profiles across donors. To identify transcriptional patterns associated with R2SN alterations, partial least squares (PLS) regression was performed using the regional R2SN t-statistic map of left-hemisphere Brainnetome regions as the imaging phenotype (Y) and regional gene expression profiles as predictors (X). Both datasets were standardized prior to analysis. PLS components capturing maximal covariance between gene expression and R2SN alterations were extracted, and the first component (PLS1) was selected for further analysis as it explained the largest proportion of variance in the imaging phenotype. Gene weights were transformed into Z-scores and compared with null distributions generated from 1000 spatial permutation (spin) tests to account for spatial autocorrelation. The spatial association between PLS1 scores and regional R2SN t-values was quantified using Pearson correlation analysis. Genes were classified as PLS1-positive or PLS1-negative according to the sign of their PLS1 weights. Genes with FDR-adjusted P < 0.05 were retained for downstream enrichment analyses, and the positive and negative PLS1 gene sets were analyzed separately. Functional enrichment analyses were conducted using Metascape, integrating Gene Ontology (GO), KEGG, and Reactome databases.59,60 To further contextualize transcriptional signatures, overlap analyses were performed between PLS1-associated genes and susceptibility genes identified in genome-wide association studies (GWAS) of neurological and psychiatric disorders. Cell-type specificity of PLS1-associated genes was assessed according to the framework proposed by Seidlitz et al, which classifies genes into seven major brain cell populations, including astrocytes, endothelial cells, microglia, excitatory neurons, inhibitory neurons, oligodendrocytes, and oligodendrocyte precursor cells (OPCs). Statistical significance was evaluated using permutation testing with FDR correction (P < 0.05). To investigate potential neurochemical substrates of R2SN alterations, spatial correlations were computed between the R2SN alteration map and normative maps of neurotransmitter receptor and transporter densities using the JuSpace toolbox, which integrates PET and SPECT datasets. Thirteen neurotransmitter-related maps were included. Spatial associations were quantified using Spearman’s rank correlation coefficients, and statistical significance was assessed using spatial permutation procedures implemented in JuSpace, with Bonferroni correction applied for multiple comparisons. Machine Learning-Based Classification Analysis To identify brain regions that most effectively distinguished T2DM-NPDR patients from controls, feature selection was performed using support vector machine–recursive feature elimination (SVM-RFE). Regional R2SN measurements based on the Brainnetome atlas were extracted, generating a 246-dimensional feature vector for each participant. Before analysis, all features were standardized using z-score normalization. SVM-RFE was implemented with a linear support vector machine, where features were ranked according to the squared magnitude of their weight coefficients and iteratively removed to obtain an optimal subset. To prevent information leakage, feature selection was conducted within a nested cross-validation framework, with SVM-RFE applied exclusively to the training data in each outer fold. Feature stability across folds was used to determine the final subset, and the top 10 brain regions were retained for subsequent analyses. The selected features were subsequently used to train five machine-learning classifiers: support vector machine (SVM), random forest (RF), logistic regression (LR), XGBoost, and LightGBM. All classifiers were evaluated within the same nested cross-validation framework, with hyperparameters optimized via grid search using only the training data. Classification performance was evaluated using accuracy, sensitivity, specificity, and the area under the receiver operating characteristic curve (AUC). Graph Neural Network Model To further evaluate the predictive capability of R2SN-derived brain networks, a graph neural network (GNN) framework was implemented for disease classification. For each participant, the individualized R2SN was modeled as a weighted undirected graph, in which nodes represented the 246 Brainnetome regions and edges represented the similarity between regional radiomic feature profiles. The adjacency matrix of the R2SN served as the graph structure input to the GNN model. A graph convolutional architecture was adopted to iteratively aggregate information from neighboring nodes and learn informative node embeddings that capture higher-order topological patterns within the brain network.53 Specifically, each graph convolutional layer updates node representations by combining information from connected regions through normalized adjacency propagation followed by nonlinear activation. The learned node embeddings were subsequently aggregated to obtain a graph-level representation for each subject. To enhance model interpretability and identify disease-relevant connections, we incorporated a learnable edge-weight masking mechanism that adaptively reweights connections during training. This masking strategy enables the model to highlight the most informative network connections contributing to classification performance. The resulting graph representation was then passed to a fully connected layer followed by a softmax classifier to predict group membership (T2DM-NPDR vs healthy controls). The model was trained using cross-entropy loss with the Adam optimizer. To evaluate model generalizability, a k-fold cross-validation strategy was applied across participants. Classification performance was assessed using accuracy, sensitivity, specificity, F1-score, and the area under the receiver operating characteristic curve (AUC). Results Demographics and Clinical Characteristics The demographic and clinical characteristics of the participants are summarized in Table 1. A total of 47 T2DM-NPDR patients and 45 healthy controls were included in the study. There were no significant differences between groups in sex distribution or age, indicating good demographic matching. In contrast, T2DM-NPDR patients exhibited significantly reduced best-corrected visual acuity in both eyes and significantly higher fasting blood glucose levels compared with healthy controls. These results confirm the expected visual impairment and metabolic dysregulation in the T2DM-NPDR group. Table 1Demographics and Visual Measurements Between Two GroupsConditionT2DM-NPDR GroupHC Groupt/χ2 valueP valueGender (male/female)18/2925/202.1010.147Age (years)54.40 ± 9.1357.38 ± 11.081.4070.163BCVA-OD0.26 ± 0.191.04 ± 0.1720.372<0.001BCVA-OS0.28 ± 0.201.01 ± 0.1619.221<0.001Fasting blood glucose (mmol/L)14.09 ± 2.834.50 ± 0.52−22.783<0.001Notes: Data are presented as mean ± standard deviation unless otherwise indicated. Sex distribution was compared using the χ2 test, and continuous variables were compared using independent-sample t-tests. Bold values indicate statistically significant group differences between the T2DM-NPDR and HC groups at P < 0.05.Abbreviations: BCVA, best-corrected visual acuity; T2DM-NPDR, type 2 diabetes mellitus with non-proliferative diabetic retinopathy; HC, healthy controls; OD, oculus dexter; OS, oculus sinister. Demographics and Visual Measurements Between Two Groups Notes: Data are presented as mean ± standard deviation unless otherwise indicated. Sex distribution was compared using the χ2 test, and continuous variables were compared using independent-sample t-tests. Bold values indicate statistically significant group differences between the T2DM-NPDR and HC groups at P < 0.05. Abbreviations: BCVA, best-corrected visual acuity; T2DM-NPDR, type 2 diabetes mellitus with non-proliferative diabetic retinopathy; HC, healthy controls; OD, oculus dexter; OS, oculus sinister. Widespread Alterations of Regional R2SN Organization in Diabetic Retinopathy Group-averaged R2SN patterns revealed broadly similar spatial organization between T2DM-NPDR patients and HC, although regional differences were observed across several cortical areas (Figure 2A and B). Region-wise two-sample t-tests identified significant alterations in R2SN values in T2DM-NPDR relative to HC, with abnormal regions distributed across both cortical and subcortical structures. Statistical significance was determined after Benjamini–Hochberg false discovery rate (FDR) correction across the 246 Brainnetome regions (q < 0.05), whereas an uncorrected threshold of P ≤ 0.01 was used for visualization purposes in Figure 2C and D. The spatial distribution of regional R2SN deviations across all Brainnetome atlas regions is shown in Figure 2E. Brain region– and network-level profiling of R2SN alterations in T2DM-NPDR. (A) Cortical surface representation of the group-averaged R2SN pattern in T2DM-NPDR patients. (B) Cortical surface representation of the mean R2SN pattern in healthy controls (HC). (C) Surface-based statistical map depicting regional R2SN differences between T2DM-NPDR and HC groups. (D) Three-dimensional brain rendering of regions showing significant R2SN alterations identified by region-wise two-sample t-tests (P ≤ 0.01). (E) Distribution of regional R2SN deviation values across Brainnetome atlas regions. (F) Network-level comparison of mean R2SN values across the seven canonical Yeo functional systems. (G) Cortical projection of significantly altered R2SN regions onto the Yeo 7 functional network atlas, highlighting the involvement of large-scale functional systems. (H) Surface mapping of the top 20 regions contributing to cognitive-related R2SN alterations. (I) Surface mapping of the top 20 structurally influential regions within the R2SN network. To further characterize large-scale system involvement, regional R2SN alterations were summarized across the seven canonical Yeo functional networks. Network-level comparisons revealed significant reductions in R2SN values in T2DM-NPDR patients across all seven functional systems (Figure 2F and Table 2), including the visual (t = −1.999, P = 0.049), somatomotor (t = −2.226, P = 0.029), dorsal attention (t = −2.528, P = 0.013), ventral attention (t = −1.996, P = 0.049), limbic (t = −2.914, P = 0.005), frontoparietal (t = −3.125, P = 0.002), and default mode networks (t = −3.062, P = 0.003). These results indicate that T2DM-NPDR-related structural network alterations extend beyond the visual system and involve multiple higher-order functional networks. Table 2Network-Level Differences in R2SN Connectivity Between Patients with T2DM-NPDR and HCConditionVisualSomatomotorDorsal AttentionVentral AttentionLimbicFrontoparietalDefaultP value0.0487080.0286210.0132740.0490870.0045320.0024150.002926t value−1.999200−2.225700−2.528100−1.995800−2.914200−3.125300−3.061800Note: p values and t values are shown for network-level comparisons between T2DM-NPDR and HC groups.Abbreviations: R2SN, regional radiomics similarity network; T2DM-NPDR, type 2 diabetes mellitus with non-proliferative diabetic retinopathy; HC, healthy controls. Network-Level Differences in R2SN Connectivity Between Patients with T2DM-NPDR and HC Note: p values and t values are shown for network-level comparisons between T2DM-NPDR and HC groups. Abbreviations: R2SN, regional radiomics similarity network; T2DM-NPDR, type 2 diabetes mellitus with non-proliferative diabetic retinopathy; HC, healthy controls. Projection of significantly altered regions onto the Yeo 7 functional network atlas further confirmed the widespread distribution of R2SN abnormalities across large-scale cortical systems (Figure 2G). Functional annotation of R2SN-associated regions highlighted prominent involvement of cognitive processes related to reward processing, motor responses, and anticipatory mechanisms (Figure 2H), including terms such as reward, anticipation, motor response, learning, and monetary incentive. Structural annotation further revealed that the most influential regions contributing to R2SN alterations were predominantly located within subcortical and basal ganglia circuits (Figure 2I), including the thalamus, putamen, globus pallidus, striatum, midbrain, and brainstem. Collectively, these findings demonstrate that T2DM-NPDR is associated with widespread structural network reorganization involving both cortical large-scale functional systems and subcortical motor–motivational circuits. Gradient Organization of the R2SN Network To further investigate the macroscale organization of the structural similarity network, gradient analysis was performed on the R2SN matrices. The principal R2SN gradient revealed a spatially continuous pattern of structural organization across the cortex in both groups (Figure 3A and B). In healthy controls, the gradient displayed a typical hierarchical organization across cortical regions, whereas T2DM-NPDR patients exhibited noticeable alterations in this gradient pattern. Direct group comparisons revealed significant regional differences in gradient values between T2DM-NPDR patients and healthy controls (Figure 3C), indicating disrupted hierarchical organization of the structural similarity network in T2DM-NPDR. Examination of the distribution of regional gradient residuals further demonstrated an overall shift in gradient organization in the T2DM-NPDR group compared with controls (Figure 3D). To determine whether these gradient alterations preferentially involved specific functional systems, gradient residuals were summarized across the seven canonical Yeo functional networks (Figure 3E and Table 3). Network-level comparisons revealed significant differences between groups in the visual (t = 2.47, P = 0.015), somatomotor (t = 2.52, P = 0.013), dorsal attention (t = 2.56, P = 0.012), and ventral attention networks (t = 2.14, P = 0.035). In contrast, no significant group differences were observed in the limbic, frontoparietal, or default mode networks (all P > 0.05; Table 3). Together, these findings suggest that T2DM-NPDR is associated with alterations in the hierarchical organization of structural similarity networks, with preferential involvement of sensory and attention-related functional systems. Gradient-based analysis of R2SNs in patients with T2DM-NPDR and healthy controls. (A) Cortical surface representation of the principal R2SN gradient in the T2DM-NPDR group, illustrating the macroscale hierarchical organization of structural similarity across cortical regions. (B) Cortical surface representation of the principal R2SN gradient in the HC group. (C) Surface-based statistical map showing regional differences in R2SN gradient values between T2DM-NPDR patients and healthy controls. Warm colors indicate higher gradient values in T2DM-NPDR relative to HC, whereas cool colors represent lower gradient values. (D) Histogram showing the distribution of regional R2SN gradient residuals in T2DM-NPDR patients (orange) and healthy controls (blue), illustrating the overall shift in gradient organization between groups. (E) Network-level comparison of R2SN gradient residuals across the seven canonical Yeo functional networks (visual, somatomotor, dorsal attention, ventral attention, limbic, frontoparietal, and default mode networks), highlighting system-level differences in gradient organization between T2DM-NPDR patients and controls. Table 3Network-Level Differences in R2SN Gradient Values Across the Yeo-7 Functional NetworksConditionVisualSomatomotorDorsal AttentionVentral AttentionLimbicFrontoparietalDefaultP value0.0154700.0134610.0123180.0353950.8852830.0755080.074788t value2.4698452.5227952.5561902.1371220.1446991.7988651.803397Note: p values and t values are shown for network-level comparisons of R2SN gradient values across the Yeo-7 functional networks.Abbreviation: R2SN, regional radiomics similarity network. Network-Level Differences in R2SN Gradient Values Across the Yeo-7 Functional Networks Note: p values and t values are shown for network-level comparisons of R2SN gradient values across the Yeo-7 functional networks. Abbreviation: R2SN, regional radiomics similarity network. Identification of Structural Network Epicenters Epicenter analysis revealed marked spatial heterogeneity in the correspondence between normative R2SN connectivity profiles and T2DM-NPDR-related R2SN abnormalities across the cortex (Figure 4A). Group-level one-sample t-tests on epicenter GOF values identified a subset of regions exhibiting significant epicenter effects following FDR correction (q < 0.05; Figure 4B). Among these regions, the highest-ranked epicenters were predominantly located in frontal regions, including the inferior frontal gyrus, middle frontal gyrus, and orbitofrontal cortex, with additional involvement of the insula and parahippocampal gyrus (Table 4 and Figure 4C). The top-ranked epicenter regions and their corresponding statistical values are summarized in Table 4. To further characterize the functional context of these epicenters, we examined the normative whole-brain functional connectivity profile of the top-ranked epicenter (Figure 4D). Although functional data were not involved in epicenter identification, this analysis provides insight into the large-scale functional systems in which the epicenter is embedded. Systems-level mapping based on the Yeo 7-network parcellation showed that the top epicenter exhibited prominent functional embedding within limbic and frontoparietal networks, with comparatively weaker associations observed in somatomotor and attention-related systems (Figure 4E). Surface-based epicenter analysis of R2SN alterations in T2DM-NPDR. (A) Cortical surface projection of epicenter goodness-of-fit (GOF) values across the whole brain, indicating the extent to which each region’s normative R2SN connectivity pattern corresponds to the spatial distribution of T2DM-NPDR-related abnormalities. (B) Surface map highlighting regions identified as significant epicenters based on group-level one-sample t-tests of GOF values, with false discovery rate (FDR) correction applied (q < 0.05). (C) Cortical visualization of the top-ranked structural-network epicenter, defined as the region with the highest group-level GOF-derived t statistic. (D) Surface-based rendering of the normative whole-brain functional connectivity profile of the top epicenter, provided for functional interpretation only and not used in epicenter identification. (E) System-level mapping of the epicenter’s normative functional connectivity distribution across the Yeo 7 large-scale functional networks. Table 4Top 10 Epicenter Regions Associated with T2DM-NPDR-Related R2SN AlterationsRankROIBrain RegionPatient_EpicenterHC_EpicenterAbsolute_EpicenterP134IFG_R_6_3−0.19−0.250.250.002238IFG_R_6_50.250.240.250.034328MFG_R_7_7−0.27−0.230.240.002443OrG_L_6_2−0.25−0.240.240.002544OrG_R_6_2−0.28−0.240.230.002627MFG_L_7_7−0.28−0.220.230.002751OrG_L_6_6−0.23−0.230.230.0028166INS_R_6_20.220.270.230.0379119PhG_L_6_60.250.210.220.0361036IFG_R_6_4−0.26−0.230.220.002Note: Epicenter values are shown for the top 10 regions associated with T2DM-NPDR-related R2SN alterations.Abbreviations: ROI, region of interest; T2DM-NPDR, type 2 diabetes mellitus with non-proliferative diabetic retinopathy; R2SN, regional radiomics similarity network; HC, healthy controls; IFG, inferior frontal gyrus; MFG, middle frontal gyrus; OrG, orbital gyrus; INS, insula; PhG, parahippocampal gyrus. Top 10 Epicenter Regions Associated with T2DM-NPDR-Related R2SN Alterations Note: Epicenter values are shown for the top 10 regions associated with T2DM-NPDR-related R2SN alterations. Abbreviations: ROI, region of interest; T2DM-NPDR, type 2 diabetes mellitus with non-proliferative diabetic retinopathy; R2SN, regional radiomics similarity network; HC, healthy controls; IFG, inferior frontal gyrus; MFG, middle frontal gyrus; OrG, orbital gyrus; INS, insula; PhG, parahippocampal gyrus. Imaging-Transcriptomic Associations of R2SN Alterations To explore potential molecular associations of T2DM-NPDR-related R2SN alterations, we performed partial least squares (PLS) regression linking regional gene expression profiles to case–control t-values derived from R2SN abnormalities. Among the first eight latent components, only the first component (PLS1) exceeded the permutation-based significance threshold based on 1000 spatial permutation tests (Figure 5A and B). Regional PLS1 scores were significantly correlated with R2SN case–control t-values (R = 0.51, P < 0.001; Figure 5C), suggesting a spatial association between cortical gene expression patterns and T2DM-NPDR-related R2SN alterations. Gene ranking based on PLS1 loadings identified representative positively weighted genes, including TNRC18, SERPINA6, BRD2, and ELMOD1, and negatively weighted genes, including IGSF8, BRD8, PRMT8, and DHPS (Figure 5D). Representative genes also showed significant regional associations with R2SN alterations (Figure 5E–H). Functional enrichment analysis of positively weighted PLS1 genes revealed significant overrepresentation of Gene Ontology (GO) categories related to chromatin regulation, transcriptional modulation, and cellular signaling processes (Figure 5I). Network-based visualization further showed clustering among enriched biological processes (Figure 5J). Transcriptomic correlates of R2SN alterations in T2DM-NPDR. (A) Null distribution of Z-scored PLS1 weights generated from 1000 spatial permutation tests. Vertical lines indicate the permutation-based significance threshold. (B) Permutation-based significance assessment of PLS components. PLS1 was significant and retained for subsequent analyses. (C) Spatial correlation between regional PLS1 scores and t-values of R2SN alterations in T2DM-NPDR (R = 0.51, P < 0.001). (D) Ranked PLS1 gene loadings, showing representative genes with positive and negative weights. (E–H) Regional associations between t-values of R2SN alterations in T2DM-NPDR and expression profiles of representative genes: TNRC18 (R = 0.31, P < 0.001), SERPINA6 (R = 0.37, P < 0.001), IGSF8 (R = −0.36, P < 0.001), and BRD8 (R = −0.45, P < 0.001). Surface maps show Z-score-normalized gene expression patterns across cortical regions. (I and J) Functional enrichment analysis of genes with positive PLS1 loadings. (I) Enriched gene Ontology terms ranked by −log10(P), with point size indicating gene count. (J) Network visualization of enriched biological processes, with node colors indicating functional categories and edges indicating semantic similarity or shared gene membership. Cell-Type–Specific Enrichment of Genes Associated with R2SN Alterations To further characterize the cellular context of T2DM-NPDR-related R2SN alterations, we examined the transcriptomic profile of negatively weighted PLS1 genes. Cell-type enrichment analysis showed that excitatory neurons (Neuro-Ex) exhibited the largest overlap with the T2DM-NPDR-related R2SN-associated gene set, followed by inhibitory neurons (Neuro-In) and astrocytes (Astro), with significant enrichment after FDR correction (P < 0.05; Figure 6A and B). Gene Ontology (GO) enrichment analysis indicated that the top-ranked biological processes among PLS1-negative genes included negative regulation of cellular component organization (GO:0051129), ribosome (GO:0042254), and cytoplasmic translation (GO:0002181), which are related to cellular organization and translational regulation (Figure 6C). Cell-type-stratified analyses further identified Neuro-Ex, Neuro-In, and Astro as the leading cellular populations across enrichment metrics (Figure 6D and E). These findings suggest that T2DM-NPDR-related R2SN alterations may be associated with cell-type-specific transcriptomic patterns, particularly involving neuronal and astrocytic signatures. Cell-type-specific characterization of R2SN-associated genes in T2DM-NPDR. (A) Spatial enrichment maps for seven major brain cell types: astrocytes (astro), endothelial cells (Endo), microglia (micro), excitatory neurons (neuro-ex), inhibitory neurons (neuro-in), oligodendrocytes (oligo), and oligodendrocyte precursor cells (OPCs). Warmer colors indicate stronger associations. (B) Overlap between R2SN-associated genes in T2DM-NPDR and cell-type-specific gene signatures. Asterisks indicate significant enrichment after false discovery rate (FDR) correction (P < 0.05). (C) Gene Ontology (GO) biological process enrichment across cell types. (D and E) Chord diagrams showing shared R2SN-associated genes among major brain cell populations. Band width indicates the magnitude of gene overlap. Spatial Associations Between T2DM-NPDR-Related R2SN Alterations and Neurochemical Systems To further characterize the neurochemical context of T2DM-NPDR-related R2SN alterations, we examined spatial associations between regional R2SN deviation maps and PET/SPECT-derived neurotransmitter receptor/transporter density maps across Brainnetome atlas regions. Spatial correlation analyses revealed significant negative associations with 5-HT1A, 5-HT2A, CB1, GABAa, and mGluR5, and significant positive associations with A4B2, D2, DAT, NMDA, and SERT (Figure 7A and B). These associations suggest that T2DM-NPDR-related R2SN alterations are spatially related to distinct normative neurotransmitter receptor and transporter distributions. All statistical significance was assessed using permutation-based null models controlling for spatial autocorrelation. Spatial correspondence between R2SN alterations in T2DM-NPDR and neurochemical systems. (A) Spatial correlations between T2DM-NPDR-related R2SN alteration maps and normative neurotransmitter receptor/transporter density maps. Significant negative correlations were observed for 5-HT1A, 5-HT2A, CB1, GABAa, and mGlur5, whereas significant positive correlations were observed for A4B2, D2, DAT, NMDA, and SERT. Statistical significance was assessed using permutation-based null models accounting for spatial autocorrelation across Brainnetome atlas regions. (B) Fisher Z-transformed Spearman correlation coefficients for each neurochemical map. Machine Learning Classification Based on R2SN Features To evaluate the discriminative ability of R2SN-derived brain features, five machine learning classifiers (SVM, Random Forest, Logistic Regression, LightGBM, and XGBoost) were trained using the top 10 features selected by SVM–recursive feature elimination. Model performance was assessed using nested cross-validation (Table 5). Among the classifiers, LightGBM achieved the highest overall performance (ACC = 0.61 ± 0.06; AUC = 0.64 ± 0.06), whereas SVM showed the highest sensitivity (0.77 ± 0.12) but lower specificity. The remaining models demonstrated comparable performance, with AUC values ranging from 0.58 to 0.62. Table 5Machine Learning Classification Performance Based on R2SN FeaturesACCAUCSensitivitySpecificityF1PrecisionSVM0.60 ± 0.040.62 ± 0.070.77 ± 0.120.43 ± 0.080.64 ± 0.060.58 ± 0.02RandomForest0.57 ± 0.080.61 ± 0.070.60 ± 0.110.55 ± 0.100.57 ± 0.090.58 ± 0.08LogisticRegression0.57 ± 0.070.58 ± 0.110.62 ± 0.290.52 ± 0.170.52 ± 0.240.46 ± 0.20LightGBM0.61 ± 0.060.64 ± 0.060.62 ± 0.140.59 ± 0.100.59 ± 0.100.59 ± 0.06XGBoost0.58 ± 0.050.62 ± 0.060.63 ± 0.100.53 ± 0.080.59 ± 0.070.58 ± 0.05Note: Values are presented as mean ± standard deviation.Abbreviations: ACC, accuracy; AUC, area under the receiver operating characteristic curve; F1, F1 score; R2SN, regional radiomics similarity network; SVM, support vector machine; RF, random forest; LR, logistic regression; LightGBM, light gradient boosting machine; XGBoost, extreme gradient boosting. Machine Learning Classification Performance Based on R2SN Features Note: Values are presented as mean ± standard deviation. Abbreviations: ACC, accuracy; AUC, area under the receiver operating characteristic curve; F1, F1 score; R2SN, regional radiomics similarity network; SVM, support vector machine; RF, random forest; LR, logistic regression; LightGBM, light gradient boosting machine; XGBoost, extreme gradient boosting. To interpret the contribution of individual features, SHAP analyses were performed (Figure 8). Across models, several brain regions consistently exhibited high feature importance. In particular, the inferior temporal gyrus (ITG_L_3_1) ranked among the most influential features and was identified as the top contributor in both RF and XGBoost models. Other prominent regions included the medial ventral occipital cortex (MVOcC_R_5_3), postcentral gyrus (PoG_L_1_1), and precentral gyrus (PrG_R_6_5). Machine learning classification and SHAP-based feature interpretation. Five machine learning models were used to classify patients with T2DM-NPDR and healthy controls based on R2SN-derived regional features: (A) LightGBM, (B) logistic regression (LR), (C) random forest (RF), (D) support vector machine (SVM), and (E) XGBoost. For each model, SHAP decision plots, heatmaps, mean absolute SHAP value plots, and beeswarm plots are presented to summarize instance-level feature contributions and global feature importance. Warmer and cooler colors in the beeswarm plots indicate higher and lower feature values, respectively. Together, these results indicate that R2SN-derived regional features contain informative patterns for distinguishing patients with T2DM-NPDR from healthy controls, with consistent contributions from distributed cortical regions across different machine learning models. Graph Neural Network-Based Classification Using R2SN Network Features To evaluate the discriminative capability of R2SN-derived network features, classification analyses were performed using matrix strength, epicenter, edge coupling, and their combined feature set (ALL). The combined feature set achieved the best overall performance across multiple evaluation metrics, including AUC, accuracy, and F1 score, compared with individual feature types (Figure 9A–C). Individual feature sets showed moderate classification performance, whereas integrating multiple network features improved overall model performance. These results suggest that complementary R2SN-derived network features provide preliminary discriminative information for distinguishing patients with T2DM-NPDR from healthy controls. Classification performance for distinguishing T2DM-NPDR from healthy controls based on R2SN-derived network features. (A) Classification performance of different feature types, including matrix strength, epicenter, edge coupling, and the combined feature set (ALL). Metrics include area under the receiver operating characteristic curve (AUC), accuracy, and F1 score; error bars indicate standard deviations across cross-validation folds. (B) Receiver operating characteristic (ROC) curves for different feature sets. (C) Heatmap summarizing AUC, accuracy, sensitivity, specificity, F1 score, and precision for each feature type. Values indicate mean performance across cross-validation folds, with standard deviations shown beneath each metric. Discussion The present study revealed widespread alterations in regional radiomics similarity network (R2SN) organization in patients with type 2 diabetes mellitus and non-proliferative diabetic retinopathy (T2DM-NPDR) compared with healthy controls. Patients with T2DM-NPDR exhibited reduced R2SN values across multiple cortical and subcortical regions, indicating disrupted structural similarity patterns at the macroscale level. Network-level analyses further demonstrated that these abnormalities were distributed across all seven canonical Yeo functional systems, including the visual, somatomotor, dorsal attention, ventral attention, limbic, frontoparietal, and default mode networks.58 Notably, relatively pronounced alterations were observed in higher-order associative systems such as the frontoparietal and default mode networks, which support executive control, internally directed cognition, and large-scale information integration.61,62 These findings suggest that brain structural network alterations in patients with T2DM-NPDR extend beyond primary visual pathways and involve distributed large-scale brain systems. Our findings are consistent with previous neuroimaging studies reporting gray matter alterations, white matter microstructural abnormalities, altered spontaneous neural activity, and cerebral perfusion changes in patients with diabetes or diabetic retinopathy.14–19 Whereas most prior studies focused on regional or modality-specific abnormalities based on conventional structural, diffusion, functional, or perfusion MRI approaches, the present study extends this literature by demonstrating that brain alterations in patients with T2DM-NPDR are reflected in individualized structural similarity networks, macroscale hierarchical organization, and spatially organized network epicenters. The integration of imaging–transcriptomic, neurochemical, and predictive modeling analyses further provides a multiscale framework for characterizing these distributed brain network alterations. Consistent with the primary clinical manifestations of diabetic retinopathy, significant alterations were also observed within the visual network.1 The retina and the brain share common embryological origins and similar neurovascular architectures, including comparable blood–tissue barriers and metabolic demands.6,63 Therefore, retinal pathology may be accompanied by alterations in central visual pathways through shared vascular, metabolic, and neurodegenerative mechanisms.63 Previous neuroimaging studies have reported abnormalities along the visual pathway in DR, including structural and functional alterations in the lateral geniculate nucleus, primary visual cortex, and higher-order visual association regions.64,65 Chronic hyperglycemia and diabetes-related microvascular dysfunction may impair neurovascular coupling, reduce cerebral perfusion, and disrupt metabolic support to visual cortical regions.66–68 In this context, the visual network findings in the present study are compatible with the retina–brain axis hypothesis. However, the involvement of multiple non-visual networks further indicates that brain structural alterations in T2DM-NPDR should be interpreted within a broader systemic neurovascular framework rather than being attributed exclusively to retinal lesions. Beyond disease-related findings, the present study also highlights the value of the R2SN framework for characterizing brain structural organization.69 Unlike conventional morphometric approaches that rely on macroscopic measures such as gray matter volume or cortical thickness, R2SN incorporates high-dimensional radiomic features derived from structural MRI.65 These features capture voxel-level intensity distributions and textural properties within brain regions, providing a richer representation of regional microstructure.70 By quantifying similarity between radiomic feature profiles across regions, R2SN enables the construction of individualized structural brain networks that may be sensitive to subtle tissue-level alterations. This property is particularly relevant for diabetes-related brain changes, where chronic metabolic dysregulation, microvascular injury, and vascular insufficiency may induce microstructural alterations before the emergence of overt macroscopic atrophy.66–68 Gradient analysis further revealed alterations in the macroscale hierarchical organization of the R2SN network in patients with T2DM-NPDR. In healthy individuals, the principal cortical gradient typically spans a continuum from unimodal sensory systems to higher-order transmodal association regions, representing a fundamental organizational principle of the human cortex.33,34 Previous neuroimaging studies have demonstrated that this unimodal–transmodal gradient captures large-scale transitions in connectivity, microstructure, and cognitive function.33,34,71 In the present study, although both patients and healthy controls exhibited broadly similar gradient architectures, patients with T2DM-NPDR showed a shift in the distribution of regional gradient values. System-level analyses indicated that these alterations were primarily localized within sensory and attention-related networks, including the visual, somatomotor, dorsal attention, and ventral attention systems. These networks occupy lower to intermediate positions along the cortical hierarchy and play critical roles in sensory integration and attentional processing.33,34,71 These findings suggest that T2DM-NPDR is associated with altered hierarchical organization of structural similarity networks, particularly within sensory–attention systems. Epicenter analysis identified candidate structural-network epicenters predominantly distributed within frontal regions, including the inferior frontal gyrus, middle frontal gyrus, and orbitofrontal cortex, with additional involvement of the insula and parahippocampal gyrus. At the systems level, these regions largely belong to the frontoparietal control network and limbic network, with additional involvement of the insular cortex associated with the salience system.72 This spatial distribution suggests that network-level abnormalities in patients with T2DM-NPDR may be preferentially associated with higher-order regulatory systems rather than being confined to primary sensory or visual pathways.72–74 Importantly, epicenter mapping identifies regions whose normative connectivity profiles resemble the spatial pattern of abnormalities and should not be interpreted as evidence of causal or temporal spread.38,39 The involvement of prefrontal regions, particularly the inferior and middle frontal gyri, suggests that structural network alterations in patients with T2DM-NPDR may affect large-scale executive control systems. The frontoparietal network plays a critical role in cognitive control, attention allocation, and adaptive behavioral regulation.73,74 Previous neuroimaging studies in diabetic populations have reported structural and functional abnormalities within prefrontal cortices, including reduced gray matter volume and altered connectivity patterns.75 Such alterations have been linked to deficits in executive function and attentional regulation frequently observed in patients with diabetes and its complications. The presence of candidate epicenter regions within these prefrontal territories suggests that the observed network abnormalities may involve higher-order cognitive control systems. In addition to the prefrontal cortex, several candidate epicenter regions were located within limbic structures, including the orbitofrontal cortex and parahippocampal gyrus. These regions are critically involved in emotional regulation, reward processing, and memory-related functions.76,77 The orbitofrontal cortex plays an important role in integrating sensory and affective information to guide decision-making, whereas the parahippocampal gyrus contributes to contextual memory and spatial processing.77–79 Alterations within these limbic regions may reflect broader effects of chronic metabolic dysregulation on emotional and mnemonic systems. Notably, cognitive decline and dementia risk are increasingly recognized in patients with diabetes and its complications, suggesting that limbic-system vulnerability may represent one potential neural context for these clinical manifestations.80,81 Furthermore, the identification of the insular cortex as one of the candidate epicenter regions highlights the potential involvement of the salience network. The insula plays a key role in integrating interoceptive signals and coordinating large-scale network switching between cognitive and affective systems.82,83 It is also closely involved in autonomic regulation and metabolic homeostasis, processes that are fundamentally disrupted in diabetes.84 Structural and functional abnormalities of the insula have been reported in several metabolic and neurovascular disorders, suggesting that this region may represent an important network hub linking systemic metabolic disturbances with large-scale brain network alterations. Taken together, the epicenter distribution identified in the present study suggests that structural network alterations in patients with T2DM-NPDR may be spatially embedded within interconnected systems responsible for cognitive control, emotional regulation, and salience detection. Rather than reflecting isolated regional abnormalities, these findings suggest that patients with T2DM-NPDR exhibit distributed alterations in higher-order brain networks, which may reflect the combined influence of retinal neurovascular injury and broader diabetes-related metabolic and vascular processes. The imaging–transcriptomic analysis further revealed that T2DM-NPDR-related R2SN alterations were spatially associated with normative molecular gradients characterized by genes involved in receptor-mediated signaling and cellular regulatory processes. Enrichment analyses highlighted biological pathways associated with GPCR signaling, MAPK cascade regulation, cellular proliferation, and tube morphogenesis, suggesting that regions showing R2SN alterations may overlap with cortical territories characterized by vascular-growth and signaling-related molecular profiles.85,86 Given that diabetic retinopathy is fundamentally characterized by microvascular dysfunction and aberrant angiogenesis, this spatial convergence may provide a potential molecular context for interpreting T2DM-NPDR-related R2SN alterations.13,87 Importantly, because the transcriptomic data were derived from normative atlases, these findings suggest that T2DM-NPDR-related structural alterations preferentially occur in cortical regions defined by specific molecular profiles rather than reflecting direct disease-induced gene expression changes.44 Cell-type–specific enrichment analyses indicated that T2DM-NPDR-related R2SN alterations were associated with a neuronal-dominant transcriptomic context. The strongest enrichment was observed in excitatory neurons, followed by inhibitory neurons and astrocytes, suggesting a possible association with molecular profiles related to cortical microcircuit organization and excitation–inhibition balance. Enrichment of biological processes related to ribosome function, cytoplasmic translation, and cellular structural organization further suggests that genes related to protein synthesis and subcellular structural maintenance may be spatially associated with R2SN alterations. The additional involvement of astrocytic signatures is consistent with the involvement of neuron–glia-related molecular signatures in the observed spatial associations, consistent with the established roles of astrocytes in metabolic support, neurotransmitter regulation, and neurovascular coupling.88,89 Complementary analyses of neurotransmitter systems revealed distinct spatial associations between T2DM-NPDR-related R2SN alterations and normative neurochemical architecture.51 Specifically, R2SN alterations showed negative spatial associations with serotonergic receptors (5-HT1A and 5-HT2A), CB1, GABAa, and mGluR5, but positive associations with dopaminergic (D2, DAT), cholinergic (A4B2), glutamatergic (NMDA), and serotonergic transporter (SERT) distributions. This bidirectional pattern suggests that T2DM-NPDR-related R2SN alterations are spatially associated with specific neuromodulatory systems. Serotonergic signaling has been implicated in neurovascular regulation and synaptic plasticity,90 while alterations in GABAergic and glutamatergic transmission have been reported in diabetes-related neural dysfunction.91,92 The observed spatial coupling provides a normative neurochemical context for interpreting T2DM-NPDR-related R2SN alterations, while further patient-specific molecular or neurochemical studies are needed to determine whether diabetes-related metabolic dysregulation directly affects these systems. Finally, machine learning and graph neural network analyses provided a complementary evaluation of the discriminative information contained in R2SN-derived features. Although several regional and network-level features contributed to the classification of patients with T2DM-NPDR and healthy controls, the overall classification performance was modest, particularly for conventional machine learning models. Therefore, these predictive findings should be interpreted as preliminary evidence that R2SN-derived features contain disease-related discriminative information, rather than as evidence for clinically useful standalone diagnostic biomarkers. The GNN framework further suggested that graph-based representations may capture higher-order topological information beyond predefined network metrics.53–55 However, larger multicenter cohorts, longitudinal data, and independent external validation are required before the translational utility of these models can be established. Taken together, the present findings suggest that T2DM-NPDR is associated with multiscale reorganization of structural brain networks. At the macroscale level, R2SN analysis revealed widespread alterations in structural similarity patterns across multiple functional systems, while gradient analysis and epicenter mapping further indicated disruptions in hierarchical organization and candidate anchor regions within higher-order integrative systems. At the molecular level, imaging–transcriptomic, cell-type enrichment, and neurotransmitter association analyses provided complementary biological context for these network alterations. Finally, predictive modeling provided preliminary support for the discriminative relevance of R2SN-derived features. Overall, these multilevel findings support the use of R2SN-based network analysis as a framework for characterizing T2DM-NPDR-related brain alterations, while further validation is needed to clarify their biological and translational significance. Limitations and Future Directions Despite the strengths of the present multimodal network framework, several limitations should be acknowledged. First, the sample size was relatively modest, which may limit the generalizability of the findings and the stability of predictive modeling results. Although nested cross-validation was employed to reduce overfitting, future studies with larger and independent cohorts are needed to further validate the robustness of the observed network patterns and their predictive utility. In addition, patients with diabetic nephropathy were excluded to reduce potential confounding effects from concomitant systemic microvascular complications and renal dysfunction on brain structure. However, this criterion may limit the generalizability of our findings to patients with diabetic retinopathy and coexisting systemic microvascular complications. Second, the imaging–transcriptomic and neurotransmitter association analyses were performed using normative molecular atlases rather than subject-specific molecular measurements. While this approach has been widely adopted to infer potential biological substrates of neuroimaging phenotypes, it does not directly reflect disease-specific gene expression or neurochemical alterations in patients with diabetic retinopathy. Future studies integrating patient-derived transcriptomic, metabolic, or molecular imaging data would provide more direct evidence regarding the biological mechanisms underlying T2DM-NPDR-related brain network alterations. Third, the present study was based on cross-sectional data, which limits the ability to determine the temporal dynamics or causal progression of brain network changes observed in patients with T2DM-NPDR. Given that type 1 and type 2 diabetes mellitus may exert distinct effects on brain structure, the present findings should be interpreted specifically in the context of patients with type 2 diabetes mellitus and non-proliferative diabetic retinopathy.93 In addition, because we did not include a diabetic control group without retinopathy, we cannot determine whether the observed brain structural network alterations are specific to NPDR or reflect broader diabetes-related metabolic and vascular effects. Longitudinal investigations including patients with T2DM without retinopathy and patients with different stages of DR will be important for clarifying disease- and stage-specific brain network alterations. Finally, although the graph neural network framework enabled topology-aware learning from the R2SN structure, the current implementation focused on structural MRI–derived networks alone. Future research incorporating multimodal neuroimaging data, such as functional connectivity, diffusion-based structural connectivity, or perfusion imaging, may further improve the characterization of brain network reorganization and enhance the predictive performance of network-based models. Taken together, addressing these limitations in future work will help refine our understanding of the neural associations between T2DM-NPDR and large-scale brain network reorganization and may contribute to the development of candidate imaging markers for this condition. Conclusion In conclusion, the present study provides a network-level characterization of brain structural reorganization associated with T2DM-NPDR using R2SNs. Our findings reveal widespread alterations in structural network organization in T2DM-NPDR, highlighting disrupted coordination among distributed brain regions beyond the classical visual system. Spatial correlation analyses further suggest that these network alterations are associated with specific neurotransmitter systems, providing potential neurochemical context for T2DM-NPDR-related brain network alterations. Moreover, machine learning models based on R2SN-derived features showed preliminary discriminative performance, suggesting that structural network patterns may provide candidate imaging markers for distinguishing patients with T2DM-NPDR from healthy controls. Collectively, these findings advance our understanding of brain network alterations associated with T2DM-NPDR from a systems-level perspective and provide new insights into the potential neurobiological context of diabetes-related brain involvement. Funding Statement This research did not receive any specific grant from funding agencies in the public, commercial, or not-for-profit sectors. Data Sharing Statement The raw data supporting the conclusions of this article will be made available. Further inquiries can be directed to the corresponding author. Ethics Statement This study was approved by the Institutional Review Board of Jiangxi Provincial Eye Hospital (IRB approval No. YLP20260046), and all research procedures adhered strictly to the Declaration of Helsinki. Participants voluntarily enrolled in the study after being fully informed of the objectives, methods, and potential risks, and provided written informed consent. Author Contributions Die Hu: Writing – original draft, Writing – review & editing, Visualization. Rui-Yang Hu: Software, Formal analysis, Methodology, Writing – review & editing. Yu-Yue Cheng: Investigation, Writing – review & editing. Fang-Liang Yu: Data curation, Resources, Writing – review & editing. Xin Huang: Conceptualization, Methodology, Supervision, Project administration, Writing – review & editing. All authors gave final approval of the version to be published; have agreed on the journal to which the article has been submitted; and agree to be accountable for all aspects of the work. Disclosure The authors declare that they have no conflicts of interest. References Associated Data Data Availability Statement The raw data supporting the conclusions of this article will be made available. Further inquiries can be directed to the corresponding author.