Brain tumors significantly impact brain networks, leading to altered functional and structural connectivity. While the tumor core and edema are easily identified via MRI, the underlying functional and diffusion signals within these regions, and their relationship to global connectivity reorganization, remain poorly understood. Previous fMRI studies have revealed altered patterns of local and global disconnections in brain tumor patients, often following functional rather than spatial proximity to the tumor. Other research has demonstrated functional abnormalities overlapping with structurally unaffected areas, and theoretical models have highlighted differences in inhibitory connections between networks with similar structural properties. However, these studies largely neglect the direct impact of the tumor on the functional signal itself and its relationship to resting-state network desynchronizations. The oscillatory nature of fMRI time series provides a valuable tool to investigate brain network interactions. The study hypothesizes that the presence of a brain tumor significantly modifies the participation of different frequencies in functional co-activations. By analyzing local and brain-wide Fourier-transformed fMRI time series, the researchers aim to quantify how deviations in frequency power propagate to alterations in resting-state network connections. Diffusion MRI (dMRI) provides insights into white matter fiber microstructure. However, existing fiber reconstruction methods struggle with the cerebrospinal fluid and gray matter abnormalities often present in tumoral tissue. The study introduces a hybrid fiber tracking pipeline designed to address this challenge. By combining a single-shell, three-tissue algorithm within the tumor area with a multi-shell algorithm in healthy tissue, the pipeline seeks to reconstruct fiber bundles within and around the tumor more accurately, minimizing the drawbacks of each individual approach. While previous studies investigating structural networks in brain tumors have yielded unclear results, this study presents a machine learning approach to predict structural rearrangement after surgery, using preoperative network data and aiming to disentangle the complex reorganization patterns in different tumor types.

Existing literature highlights altered functional connectivity in brain tumor patients, often correlated with functional distance from the tumor rather than purely spatial proximity (Stoecklein et al., 2020; Nenning et al., 2020; Silvestri et al., 2022). These studies demonstrate widespread functional disconnections, even in areas seemingly unaffected structurally. Theoretical modeling suggests differences in inhibitory connections, even when structural features are similar (Deco & Corbetta, 2011; Deco et al., 2009). However, a gap remains in understanding how tumors directly modify functional signals and their relationship to resting-state network desynchronizations. The oscillatory nature of fMRI signals, offering insight into cognitive processes (Bullmore & Bassett, 2011), is underutilized in this context. Similarly, while dMRI offers promising insights into brain tumor microstructure (Nilsson et al., 2018), current methods struggle with the complex tissue composition within and around tumors. Previous research employing dMRI in brain tumors has shown mixed results regarding tumor-dependent differences in network topology, with some studies finding minimal alteration (Yu et al., 2016; Fekonja et al., 2022). The connection between structural and functional connectivity remains a crucial open question (Suárez et al., 2020; Crimi et al., 2021; Faskowitz et al., 2022), particularly regarding plasticity and recovery after tumor resection. Machine learning approaches, though showing promise in brain network analysis (Hamilton et al., 2017; Sarwar et al., 2021), face challenges with complex tumor-related network topologies (Dehmamy et al., 2019).

This study used a multi-modal MRI approach, combining fMRI and dMRI, to assess functional and structural changes in brain networks of patients with brain tumors. The dataset consisted of pre- and post-surgical MRI scans from 36 subjects (11 healthy controls, 14 meningioma patients, and 11 glioma patients). Data from 28 subjects were available post-surgery. **Functional MRI (fMRI) Analysis:** Resting-state fMRI data were analyzed to assess functional connectivity within the Default Mode Network (DMN). The study segmented the tumor, including edema, necrotic tissue, and enhancing components. Functional signals were extracted from both the DMN regions and the tumor. A Dynamics Alteration Score (DAS) was developed based on the power distribution of Fourier transformed BOLD signals to quantify changes in oscillatory dynamics, comparing patients to controls. The Richness score was also used to measure the complexity of the functional networks. Spatial relationships (distance and overlap) between tumor and DMN were analyzed. **Diffusion MRI (dMRI) Analysis:** A hybrid tractography pipeline was developed to reconstruct white matter fiber tracts within and around the tumor. This pipeline combined a single-shell, three-tissue constrained spherical deconvolution (SS3T) algorithm within the tumor mask with a multi-shell, multi-tissue (MSMT) approach in healthy tissue. This methodology is thought to improve fiber bundle reconstruction in lesioned areas. The resulting structural connectomes were analyzed to evaluate alterations in connectivity. **Machine Learning:** A fully connected network (FCNET) was used to predict post-surgical structural connectivity based on pre-surgical data, considering the impact of tumor resection. The model incorporated an anatomical prior derived from healthy subjects to guide predictions. The FCNET was compared to a Huber regressor and a null model, using metrics such as Mean Squared Error (MSE), Mean Absolute Error (MAE), Pearson Correlation Coefficient (PCC), Cosine Similarity (CS), Kullback-Leiber (KL) divergence, and Jensen-Shannon (JS) divergence to assess model performance. Leave-one-out cross-validation was employed. **Statistical Analysis:** Various statistical tests, including two-tailed and one-tailed t-tests, U-tests, permutation tests, ANOVA and Kruskal-Wallis tests were used to compare groups and analyze correlations between variables. The normality of data distributions was also assessed.

The study yielded several key findings: **Functional Connectivity:** Analysis of fMRI data revealed alterations in the power distribution and cumulative power distributions of BOLD oscillations within the DMN of brain tumor patients. The DAS, designed to capture changes in oscillation dynamics, showed significant correlations with node similarity within the DMN and changes in network complexity, indicating a relationship between altered BOLD dynamics and functional network reorganization. Interestingly, the origin of these desynchronizations was not solely explained by the spatial proximity of the tumor to the DMN. The study found a strong correlation between alterations in the DMN and in the BOLD signals within the tumors themselves, suggesting a possible causal link between local tumor activity and global network effects. **Structural Connectivity:** The hybrid fiber tracking pipeline successfully reconstructed white matter fiber tracts in the presence of tumors, recovering pathways that were otherwise truncated by conventional methods. The results suggest the hybrid approach is at least equivalent to state-of-the-art methods, especially for larger tumors. Scale-free properties of the reconstructed networks were preserved, indicating that the methodology did not introduce substantial artificial changes in connectivity. **Machine Learning Predictions:** The FCNET significantly outperformed benchmark models (Huber regressor and null model) in predicting post-surgical structural connectivity. The model's accuracy was influenced by tumor size, with smaller tumors leading to more accurate predictions. While the model showed some sensitivity to tumor grade and location, it was less sensitive to tumor histology and periventricular status. The FCNET generated networks that exhibited similar biological topological characteristics to real networks, including lognormal weight probability distributions.

This study provides crucial insights into functional and structural brain network reorganization in brain tumor patients. The findings demonstrate a strong relationship between altered BOLD dynamics within the tumor and global DMN desynchronization, suggesting that local tumor activity can propagate to disrupt large-scale brain networks. The development of a novel fiber tracking pipeline enhances the reconstruction of white matter tracts in challenging tumoral regions, improving the accuracy of structural connectome mapping. The machine learning model provides a robust tool to predict post-surgical structural reorganization, taking into account tumor size, grade, and location while identifying confounding effects.

This study demonstrates the significance of analyzing fMRI and dMRI signals within the tumoral region to understand brain network reorganization in brain tumor patients. The hybrid tractography pipeline improves structural network reconstruction, and the machine-learning model offers effective prediction of post-surgical connectivity. These findings underscore the dynamic nature of functional and structural changes in brain tumors and highlight the potential of combined neuroimaging and machine learning techniques for improving diagnosis, prognosis, and surgical planning. Future research should focus on larger datasets to validate findings and explore the predictive potential of the DAS score regarding cognitive impairment and survival.

The primary limitation of this study is the relatively small sample size, potentially impacting the generalizability of the findings. The heterogeneity of brain tumors and their location further complicates the analysis, necessitating caution in interpreting the results. Further research using larger, multi-site datasets is essential to validate these results and investigate the influence of individual patient characteristics more thoroughly. The reliance on a specific machine learning model may also limit the broader applicability of the predictive approach.