Abstract
Background
The mechanism by which activity in the brain relates to its anatomical structure is a central topic of research in neuroscience. In recent years, this relation has been studied in the context of network science, where anatomical white-matter connections are captured in Structural Connectivity (SC) and statistical dependence between grey-matter activity in Functional Connectivity (FC). The relation between SC and FC is poorly understood at an individual level, especially in glioma patients. Developing methods to predict FC from SC in these patients could improve our understanding of how gliomas affect brain function, and consequently cognition. Furthermore, a model accurate at individual patient level could aid in clinical decision-making by identifying eloquent areas and connections in individuals.
Material and methods
Building on earlier work predicting FC from SC in healthy subjects, we trained a deep learning model on tractography and resting-state fMRI data of 288 glioma patients to predict FC at an individual level. Since our sample size is modest in the context of deep learning, we attempted to improve accuracy by pre-training the model on a large set of healthy subjects. This way, the model should learn baseline patterns in SC and FC before learning the specific patterns associated with glioma patients. We used data from 1052 subjects of the Human Connectome Project and 633 subjects of the Cambridge Centre for Ageing and Neuroscience. Finally, we analyzed the effect of several factors on prediction accuracy: tumor volume, location and grade and patient age.
Results
Training on only the patient set, the prediction model achieved an accuracy, measured as correlation between predicted and empirical FC, of 0.53 ± 0.03 at individual patient level. This is comparable with accuracy achieved in studies on healthy subjects. Pre-trained models did not improve accuracy. The model had more difficulty predicting FC in patients with a glioma in the insular region (p = 0.0014). No other factors were associated with accuracy.
Conclusion
Although further research is required to investigate the utility of FC predictions made by our model, our work is a first step towards predicting activity from brain structure in glioma patients at an individual level, since earlier research in patient populations has focused on group-level effects. We observed that predictions are less accurate in patients with a glioma in the insular region, which indicates that the relation between the structural and functional organization of the brain in this group is different from that in patients with tumors in other locations. Improving our methods to achieve higher accuracy could provide us, as a next step, with a tool to predict functional consequences of resective surgery on glioma patients, and consequently improve decision making by more accurately identifying eloquent regions at an individual level.
The mechanism by which activity in the brain relates to its anatomical structure is a central topic of research in neuroscience. In recent years, this relation has been studied in the context of network science, where anatomical white-matter connections are captured in Structural Connectivity (SC) and statistical dependence between grey-matter activity in Functional Connectivity (FC). The relation between SC and FC is poorly understood at an individual level, especially in glioma patients. Developing methods to predict FC from SC in these patients could improve our understanding of how gliomas affect brain function, and consequently cognition. Furthermore, a model accurate at individual patient level could aid in clinical decision-making by identifying eloquent areas and connections in individuals.
Material and methods
Building on earlier work predicting FC from SC in healthy subjects, we trained a deep learning model on tractography and resting-state fMRI data of 288 glioma patients to predict FC at an individual level. Since our sample size is modest in the context of deep learning, we attempted to improve accuracy by pre-training the model on a large set of healthy subjects. This way, the model should learn baseline patterns in SC and FC before learning the specific patterns associated with glioma patients. We used data from 1052 subjects of the Human Connectome Project and 633 subjects of the Cambridge Centre for Ageing and Neuroscience. Finally, we analyzed the effect of several factors on prediction accuracy: tumor volume, location and grade and patient age.
Results
Training on only the patient set, the prediction model achieved an accuracy, measured as correlation between predicted and empirical FC, of 0.53 ± 0.03 at individual patient level. This is comparable with accuracy achieved in studies on healthy subjects. Pre-trained models did not improve accuracy. The model had more difficulty predicting FC in patients with a glioma in the insular region (p = 0.0014). No other factors were associated with accuracy.
Conclusion
Although further research is required to investigate the utility of FC predictions made by our model, our work is a first step towards predicting activity from brain structure in glioma patients at an individual level, since earlier research in patient populations has focused on group-level effects. We observed that predictions are less accurate in patients with a glioma in the insular region, which indicates that the relation between the structural and functional organization of the brain in this group is different from that in patients with tumors in other locations. Improving our methods to achieve higher accuracy could provide us, as a next step, with a tool to predict functional consequences of resective surgery on glioma patients, and consequently improve decision making by more accurately identifying eloquent regions at an individual level.
| Original language | English |
|---|---|
| Journal | Neuro-Oncology |
| Volume | 25 |
| Issue number | Supplement_2 |
| DOIs | |
| Publication status | Published - 2023 |