Predicting cognitive functioning for patients with a high-grade glioma: Evaluating different representations of tumor location in a common space
Academic Background
It is widely recognized that the cognitive function of patients with high-grade glioma is affected by the location and volume of the tumor. However, research on how to accurately predict individual patients’ cognitive function for personalized treatment decisions before and after surgery remains limited. Currently, most studies focus on explaining the effects of different tumor locations on cognitive function but do not explore whether these location representations can be used for actual prediction. Moreover, most methods used are based on population-averaged brain atlases, which may not accurately reflect individual differences. This study aims to explore whether different tumor location representation methods, including popular population-averaged brain atlases, randomly generated atlases, and Principal Component Analysis (PCA)-based representations, can effectively predict the cognitive function of unseen patients.
Research Source
The paper is titled “Predicting cognitive functioning for patients with a high-grade glioma: evaluating different representations of tumor location in a common space.” The main authors include S. M. Boelders, W. De Baene, E. Postma, K. Gehring, and L. L. Ong, who are from the Cognitive Science and Artificial Intelligence Department and the Neuropsychology Department at Tilburg University. The paper was published on May 31, 2024, in the journal Neuroinformatics.
Research Process
Study Design and Participants
This study included 246 patients with high-grade glioma (WHO grade 3 or 4) who underwent surgery at Elisabeth-Tweesteden Hospital between 2010 and 2019 and had preoperative cognitive screening. Participants ranged in age from 18 to 81 and underwent standardized interviews to obtain demographic variables such as age, gender, and educational background. Cognitive screening used the computerized CNS Vital Signs (CNS VS) test battery, including eight cognitive tests.
Neuroimaging Processing and Segmentation
Each patient underwent preoperative MRI scans, including T1, T1 contrast-enhanced, T2, and FLAIR sequences. The scan results were registered to the MNI space using the REGALADIN tool in the LINDA package and underwent skull stripping using HD-BET. Tumor contrast-enhanced areas were segmented using a convolutional neural network (U-Net architecture) and manually corrected.
Tumor Location Representation Methods
The study reduced each high-dimensional voxel segmentation data to 39 different low-dimensional representations, including 13 population-averaged atlases, 13 random atlases, and 13 PCA-based representations. Each representation method is as follows:
- Population-Averaged Atlas Representation: Calculating the overlap percentage of each region by overlapping the tumor segmentation with the population-averaged atlas.
- Random Atlas Representation: Generating by randomly selecting seed voxelsusing a grassfire algorithm.
- PCA Representation: Using PCA on voxel tumor segmentations to derive a small number of uncorrelated features.
Predicting Cognitive Function
An ElasticNet model was used to predict the patients’ cognitive test results and compare the predictive performance of each representation method. Leave-one-out cross-validation was used to evaluate model performance, and results were compared with a baseline model that only used tumor volume as a predictor.
Research Results
Baseline Model Performance
The baseline model used only tumor volume as a predictor, with the variance explained for different cognitive tests ranging from less than 1% to 9.6%.
Performance of Different Representation Methods
Overall, the performance of various representation methods was similar, and no representation method significantly outperformed the baseline model. PCA-based representations slightly outperformed in most cases but with minimal differences. Specifically, the PCA method outperformed random atlases in three tests and outperformed population-averaged atlases in seven tests.
Key Findings
- Individual Functional Area Differences: Population-averaged atlases did not significantly improve predictive performance in most cases, possibly due to the mass effect of tumors decreasing the accuracy of the atlases.
- Potential of PCA Method: While differences were small, PCA-based methods performed slightly better in most cases, suggesting it might have more predictive potential than traditional atlases.
- Poor Performance of High-Dimensional Representations: High-dimensional representations (such as atlases with more than 100 regions) performed poorly in prediction, consistent with the theory of the “curse of dimensionality.”
Conclusion and Application Value
The results indicate that although different tumor location representation methods have theoretical predictive value, they do not significantly outperform the baseline model using only tumor volume in practice. This suggests that commonly used tumor location representation methods may have limitations when predicting the cognitive function of individual patients. It underscores the necessity of developing and validating new methods that can accommodate individual differences and the influence of tumor lesions. PCA-based methods showed some potential but require further research to validate their generalizability in different datasets.
The study highlights the importance of developing more precise individual brain region segmentation methods in the field of neuroscience while considering individual differences in brain structure and function. This could significantly improve the accuracy of cognitive function predictions. For researchers dedicated to predicting patients’ cognitive function, PCA-based methods might be a viable option, although understanding the biological significance of PCA features may present challenges.