Functional and structural reorganization in brain tumors: a machine learning approach using desynchronized functional oscillations

The study investigates how brain tumors alter functional and structural brain organization, focusing on signals within the lesion that are typically excluded from analyses. The central questions are: (1) how the presence of a tumor modifies the oscillatory content of BOLD signals and how such changes relate to desynchronization and reorganization of resting-state networks (specifically the Default Mode Network, DMN); (2) how to reconstruct white-matter connections within and around tumoral regions despite diffusion signal contamination; and (3) whether post-surgical structural connectivity can be predicted from preoperative networks using machine learning guided by healthy anatomical priors. The work is motivated by prior observations of altered functional connectivity patterns in tumor patients that do not strictly follow spatial proximity to lesions, and by unresolved issues linking structure to function in the context of pathology. Understanding these mechanisms could inform surgical planning and prognostication of network-level recovery.

Prior fMRI work in brain tumor patients shows altered local and global functional connectivity that follows functional rather than spatial distance to tumors, with abnormalities overlapping structurally unaffected areas. Large-scale models indicate differences in inhibitory connections despite similar structural features. In diffusion MRI, within-tumor tractography is challenged by CSF contamination and gray matter abnormalities; while intra-lesion tracking methods exist, their adoption is limited due to acquisition and modeling compromises. Studies of structural networks have reported inconsistent tumor-dependent differences, with ipsilesional alterations and potential post-surgical normalization. Spectral Graph Theory and related approaches posit structural constraints underpin functional activity and behavior, yet structure-function coupling remains incompletely understood. Machine learning on graphs (e.g., GNNs) faces difficulties aggregating long-range information in complex topologies and demands large datasets; modular designs can help but become data-hungry. Fully connected layers, combined with anatomical priors, offer a simpler alternative for generative modeling in data-scarce clinical contexts.

Data: Resting-state fMRI and multi-shell diffusion MRI from brain tumor patients and controls. For functional analyses, 25 patients and 21 controls were available. For structural learning, 19 pre-/post-surgery connectome pairs were used in Leave-One-Out cross-validation. Functional analyses: BOLD time series were extracted from tumor masks (including edema) and from 41 DMN regions (Gordon parcellation). Signals were Fourier-transformed to compare power spectra independent of phase. Total power and binned power distributions were computed, along with cumulative power (CP). A Dynamics Alteration Score (DAS) was defined as the signed area between cumulative power curves of patient and healthy signals, with positive DAS indicating relatively slower dynamics in the patient. Network-level metrics included pairwise correlation-based functional connectivity, node-wise similarity to healthy templates, and a Richness score quantifying deviation of the correlation distribution from uniform. Structural tractography: A hybrid pipeline combined two constrained spherical deconvolution approaches: (1) multi-shell multi-tissue (MSMT-CSD) with anatomically constrained tractography (ACT) in non-lesioned tissue to reconstruct "healthy" fibers; (2) single-shell 3-tissue (SS3T-CSD) run only inside the lesion (using the highest b-shell, b=2800 s/mm²) to recover intra-/peri-tumoral FODs without ACT constraints where anatomy is unreliable. FOD images from MSMT and SS3T were merged (non-overlapping), tractography used iFOD2 with dynamic seeding; within-lesion seeding counts were calibrated by the number of streamlines overlapping the lesion projected onto healthy controls. SIFT2 provided streamline weights. Structural connectivity matrices were built from weighted streamline counts (log(1+w) normalization). Lesion and healthy matrices were merged greedily (element-wise max) to form preoperative lesioned networks. Controls and post-op patients used standard MSMT pipelines. Machine learning prediction: Goal was to predict post-surgical structural connectivity from preoperative networks. A one-hidden-layer fully connected neural network (FCNET) was trained with an anatomical prior (from healthy cohorts) in a Bayesian-inspired framework that weighted outputs by prior plausibility. Models compared: FCNET, Huber Regressor, and an untrained Null linear generator; Huber and Null outputs were likewise weighted by the same anatomical prior. Evaluation used LOOCV (19 folds), with internal train/validation splits. Metrics: MSE, MAE, Pearson correlation coefficient (PCC), cosine similarity (CS), and topological KL and Jensen–Shannon (JS) divergences between predicted and ground-truth weight distributions. Additional analyses assessed subject-specificity, normality of z-scores, sensitivity to tumor features (size, histology, grade, location, periventricular). Imaging preprocessing followed standard pipelines (fMRIPrep/XCP-D for fMRI; MRtrix3-based corrections for dMRI), with TR harmonization via zero-padding where needed.

Functional domain: - Tumor and DMN signals: Qualitative similarity of raw BOLD signals, but frequency-domain alterations present in some patients. No consistent differences in signal complexity alone. - DAS and network reorganization: Node-wise similarity of patients’ DMN to healthy templates was significantly anti-correlated with the magnitude of DAS (r = −0.506, df = 23, p = 0.01). An inverted U-shape indicated that departures in either direction (slower or faster) reduced similarity. DAS correlated with changes in functional complexity (Richness): r = 0.514, df = 23, p = 0.009; also significant for absolute alterations (r = 0.413, p = 0.04). - Spatial relations: DAS showed no correlation with mean Euclidean distance (p = 0.583) or overlap (p = 0.29) between tumor and DMN centroids. - Tumor-DMN linkage: DAS within tumor and within the DMN of the same patient were highly correlated (r = 0.696, df = 23, p < 0.001), linking intra-lesion oscillatory changes with distributed network desynchronization. - Power vs dynamics: Intra-lesion power was elevated relative to healthy (P > 0, p = 0.031), but dynamics (|DAS|) showed stronger deviations (p < 0.001). There was no relationship between intensity changes and oscillation alterations (r = −0.113, p = 0.588). - Tumor features: Periventricular tumors tended to have higher positive DAS (slower dynamics), not significant (N=5, p = 0.31); logistic regression yielded R² ≈ 0.8 but underpowered. Structural domain (tractography and prediction): - Hybrid tractography: MSMT alone overdamped lesion diffusion signal and ACT caused premature streamline termination. The hybrid MSMT+SS3T approach with relaxed peritumoral constraints reconstructed known bundles (e.g., corticospinal tract, SLF) that were truncated otherwise, integrating seamlessly with non-lesioned fibers. - Topology of reconstructed networks: Weighted degree distributions showed scale-free properties in asymptotic regimes (α ∈ [2,3], KS distances 0.15–0.22). Differences between hybrid and standard MSMT pipelines were comparable for small tumors and increased with tumor size; tumor size was the strongest predictor of differences (R² = 0.439, p = 0.006; size p = 0.005). - Predictive modeling: FCNET outperformed Null on all metrics (p < 0.001) and outperformed Huber on numerical/similarity metrics (MSE, MAE p < 0.05; PCC, CS p ≈ 0.05), but not on topological divergences (KL, JS p > 0.4). Both FCNET and Huber reduced KL and JS vs Null (p < 0.001). FCNET produced a small fraction of negative edges (<25%, magnitudes 0 to −0.5 in log scale), which would be thresholded anatomically. - Subject-specificity/robustness: Metrics’ z-scores were near-normal; ~65% within ±10, ~95% within ±20; retraining with different initializations changed scores for all but 2 subjects, which were not consistent outliers. Q–Q plots showed r > 0.9, p < 0.01 for normality. - Sensitivity to tumor features: Overall correlation of PCC with size not significant; excluding the three largest tumors (≥60 cm³) revealed a significant negative correlation (r = −0.336, p = 0.04), stronger upon excluding the fourth largest (r = −0.451, p < 0.01). Dichotomizing by median size (12.95 cm³): small tumors had higher PCC (0.825 ± 0.007) than large (0.792 ± 0.013), p = 0.03. Histology (meningioma vs glioma) and periventricular involvement did not significantly affect PCC; lower grade trended better than higher grade; frontal tumors showed slightly higher PCC than temporal/parietal (p = 0.12). Predicted networks reproduced biological lognormal weight distributions and topological properties without explicit topological loss terms.

Findings indicate that tumor-induced alterations in the oscillatory content of BOLD signals are tightly linked to reorganization of the DMN: larger deviations in frequency content (DAS) are associated with reduced node-wise similarity to healthy networks and with changes in functional complexity. The lack of association between DAS and spatial overlap or distance to DMN suggests that desynchronization is not purely a function of spatial proximity; nevertheless, intra-tumor dynamics strongly track DMN alterations, consistent with a scenario in which local tumor-driven changes propagate through structural pathways to affect distributed networks. Slower BOLD dynamics (DAS > 0) associated with increased complexity may reflect more temporally coherent activity patterns, whereas faster dynamics may degrade coordinated interactions and flatten correlation distributions. Periventricular tumors showed a trend toward slower dynamics, potentially related to CSF/lymphatic fluid effects within lesions, though larger cohorts are needed. Structurally, a hybrid tractography strategy that leverages MSMT in healthy tissue and SS3T within lesions while relaxing ACT constraints perilesionally can recover plausible intra-/peri-tumoral tracts otherwise truncated, mitigating artificial disconnections that could bias network analyses and surgical planning. The predictive FCNET model, guided by healthy anatomical priors, captures post-surgical network characteristics and preserves biologically plausible topologies, outperforming benchmarks on numerical and similarity metrics. Performance degrades with larger tumors, plausibly due to disruption of long-range, metabolically costly connections and the emergence of non-normative, indirect rewiring patterns that are harder to model. Tumor grade appears more relevant than histology for prediction sensitivity, and frontal locations may afford slightly more normalized rewiring patterns captured by the model. Importantly, both functional and structural results underscore complex, tumor-dependent plasticity and emphasize inclusion of intra-lesion signals in analyses.

The study provides evidence that brain tumors induce measurable changes in the frequency content of BOLD signals that relate to DMN reorganization, with intra-lesion dynamics tightly linked to distributed network alterations. A hybrid tractography pipeline effectively reconstructs intra- and peri-tumoral fibers while preserving anatomical information elsewhere, reducing artificial disconnections. A simple, fully connected generative model, regularized by healthy anatomical priors, predicts post-surgical structural connectivity with strong numerical and similarity performance and maintains biologically plausible topologies. Together, these contributions highlight the importance of incorporating lesion signals into both functional and structural analyses when planning and evaluating neurosurgical interventions. Future work should include larger, multi-site cohorts; integrate causal modeling to clarify directionality between tumor and network changes; test graph-based deep architectures with topological guidance adapted to small datasets; and link imaging-derived metrics (e.g., DAS) to cognitive outcomes and survival, including molecular tumor profiles.

Primary limitations include small sample size and absence of an external validation cohort, limiting generalizability and statistical power. Heterogeneity across patients (tumor size, location, grade/histology, timing between scans) introduces confounds despite subgroup analyses. Functional TR differences required zero-padding harmonization. The tractography validation lacks a definitive ground truth; anatomical constraints were relaxed perilesionally, which may introduce biases despite SIFT2 weighting. The predictive model, constrained to a one-hidden-layer architecture due to data scarcity, may underfit complex, non-normative rewiring in large or highly infiltrative tumors.