Abstract
BACKGROUND
Predicting function from structure is central to optimizing the onco-functional balance in neurosurgical planning: if we can predict changes in function due to changes in structure, we may be able to predict the functional outcomes of surgery. In recent years, this challenge has been formulated in terms of predicting Functional Connectivity (FC) from Structural Connectivity (SC), both derived from advanced MRI. Several recent studies claim high individual-level accuracy in predicting FC from SC on healthy subjects. We investigated whether these approaches could be extended to the prediction of FC from SC in glioma patients.
MATERIAL AND METHODS
Three recent studies with promising results were considered for adaptation. To ensure correct implementation, we reproduced their methods on 1000 healthy subjects of the Human Connectome Project (HCP), a data set used in all three studies. We also retrospectively included 242 glioma patients from the Elisabeth-TweeSteden hospital (Tilburg, The Netherlands) to investigate if the studies’ methods could be translated to this population. SC was derived from tractography on diffusion-weighted MRI, while FC was derived from resting-state functional MRI. Performance was calculated as the average individual-level correlation between FC predicted from SC and FC derived from fMRI, and assessed with 10-fold cross-validation, using 90% of the data for training and 10% for validation. To compare performance on the two different data sets in a meaningful way, we constructed a baseline predictor that calculates the group average fMRI-derived FC of the training set and uses this single matrix as predictor for the validation set.
RESULTS
We replicated the prediction performance of 0.71 reported by the first of the three studies using HCP data. On the glioma patient data, we achieved a performance of 0.43. However, we found that for both data sets, the group average predictor outperformed the prediction model presented by the study, achieving a performance of 0.84 and 0.56 respectively. Subsequently, we found that the group average predictor matched the prediction performance reported by the other two studies.
CONCLUSION
None of the three studies improved upon the basic group average predictor, indicating that none of the predictive models represent a meaningful individual mapping from SC to FC. We consider it likely that the construction of SC or FC is currently not accurate enough at individual level to capture meaningful variation in individuals, either due to excessive noise in resting-state fMRI or due to inadequate accuracy in individual-level parcellations. In conclusion, a failure of promising methods to work on clinical populations may not be caused by population-specific difficulties, but by subtle problems with the methods themselves, only revealed upon careful scrutiny of the implications of the presented quantitative results.
Predicting function from structure is central to optimizing the onco-functional balance in neurosurgical planning: if we can predict changes in function due to changes in structure, we may be able to predict the functional outcomes of surgery. In recent years, this challenge has been formulated in terms of predicting Functional Connectivity (FC) from Structural Connectivity (SC), both derived from advanced MRI. Several recent studies claim high individual-level accuracy in predicting FC from SC on healthy subjects. We investigated whether these approaches could be extended to the prediction of FC from SC in glioma patients.
MATERIAL AND METHODS
Three recent studies with promising results were considered for adaptation. To ensure correct implementation, we reproduced their methods on 1000 healthy subjects of the Human Connectome Project (HCP), a data set used in all three studies. We also retrospectively included 242 glioma patients from the Elisabeth-TweeSteden hospital (Tilburg, The Netherlands) to investigate if the studies’ methods could be translated to this population. SC was derived from tractography on diffusion-weighted MRI, while FC was derived from resting-state functional MRI. Performance was calculated as the average individual-level correlation between FC predicted from SC and FC derived from fMRI, and assessed with 10-fold cross-validation, using 90% of the data for training and 10% for validation. To compare performance on the two different data sets in a meaningful way, we constructed a baseline predictor that calculates the group average fMRI-derived FC of the training set and uses this single matrix as predictor for the validation set.
RESULTS
We replicated the prediction performance of 0.71 reported by the first of the three studies using HCP data. On the glioma patient data, we achieved a performance of 0.43. However, we found that for both data sets, the group average predictor outperformed the prediction model presented by the study, achieving a performance of 0.84 and 0.56 respectively. Subsequently, we found that the group average predictor matched the prediction performance reported by the other two studies.
CONCLUSION
None of the three studies improved upon the basic group average predictor, indicating that none of the predictive models represent a meaningful individual mapping from SC to FC. We consider it likely that the construction of SC or FC is currently not accurate enough at individual level to capture meaningful variation in individuals, either due to excessive noise in resting-state fMRI or due to inadequate accuracy in individual-level parcellations. In conclusion, a failure of promising methods to work on clinical populations may not be caused by population-specific difficulties, but by subtle problems with the methods themselves, only revealed upon careful scrutiny of the implications of the presented quantitative results.
Original language | English |
---|---|
Pages (from-to) | V62-V63 |
Journal | Neuro-Oncology |
Volume | 26 |
Issue number | Supplement_5 |
DOIs | |
Publication status | Published - 2024 |