Brain age based signatures as a tool for differentiating neurological conditions
Poster No:
1187
Submission Type:
Abstract Submission
Authors:
Rafael Navarro-González1, James Cole2,3, Rodrigo De Luis-García1, Santiago Aja-Fernández1
Institutions:
1Laboratorio de Procesado de Imagen, UVA, Valladolid, España, 2Centre for Medical Image Computing, UCL, London, United Kingdom, 3Dementia Research Centre, UCL, London, United Kingdom
First Author:
Co-Author(s):
James Cole, PhD
Centre for Medical Image Computing, UCL|Dementia Research Centre, UCL
London, United Kingdom|London, United Kingdom
Centre for Medical Image Computing, UCL|Dementia Research Centre, UCL
London, United Kingdom|London, United Kingdom
Introduction:
Brain age, a measurment derived aplying machine learning to brain MRI images for age prediction, has demonstrated being valuable in characterizing diseases like Alzheimer's and multiple sclerosis [2-3]. However, it is a reductionist approach which simplifies all the information of an image into one number, the predicted age. Further information about different diseases could be captured if the image is transformed into a brain age signature which could be later used as the input of a classification task among different diseases. In this work, we study this approach by training a brain age model on the output of the brain volume segmentation of a manifold of T1w MRI images from healthy adults. The selected features for the brain age task were then used for a classification task between healthy adults, chronic schizophrenic and chronic migraine patients.
Methods:
A total of 2850 T1w MRI images from different free access datasets and 191 local data acquisitions have been used for the brain age task, while 65 schizophrenic patients, 74 chronic migraine patients, from our own institution and 94 healthy individuals, subset of the healthy patients selected for the study, ages between 18 to 60, were selected for the classification task (See Table 1). FastSurfer [4], a deep learning whole brain segmentation method trained on the Desikan-Killiany atlas [5], was utilized to segment the dataset acquisitions and feature extraction. A total of 624 morphological features were calculated over the segmented regions of interest (ROIs) encompassing area, volume, curvature and thickness of cortical and subcortical brain ROIs.
The dataset was split into training and testing (80/20 ratio). Training involved a three-fold cross-validation with a Multilayer perceptron (MLP) regressor. In each fold, outliers (defined as 2.5 and 97.5 quantile values) were removed, data was normalized (range -1, 1) and feature selection was performed. Three feature sets (25, 50, and 100 features) were created using a two-step method: initial filtering for the top 20% of features by mutual information with age, followed by refinement with a forward feature selection algorithm using Gaussian mixture models. The same process was used for an MLP classifier. Hyperparameters for both MLPs were determined during validation, and models were retrained on the entire training cohort and tested on test data. Brain age prediction was evaluated using mean absolute error (MAE), Pearson's correlation (r), and R-squared (R²), while classification performance was measured by Precision, Recall, Area Under the Curve (AUC), and Matthew's Correlation Coefficient (MCC). Following the same scheme, another MLP was trained in selecting features with the disease label as the target. A comparison between the brain age-based classification and this procedure was performed.
The dataset was split into training and testing (80/20 ratio). Training involved a three-fold cross-validation with a Multilayer perceptron (MLP) regressor. In each fold, outliers (defined as 2.5 and 97.5 quantile values) were removed, data was normalized (range -1, 1) and feature selection was performed. Three feature sets (25, 50, and 100 features) were created using a two-step method: initial filtering for the top 20% of features by mutual information with age, followed by refinement with a forward feature selection algorithm using Gaussian mixture models. The same process was used for an MLP classifier. Hyperparameters for both MLPs were determined during validation, and models were retrained on the entire training cohort and tested on test data. Brain age prediction was evaluated using mean absolute error (MAE), Pearson's correlation (r), and R-squared (R²), while classification performance was measured by Precision, Recall, Area Under the Curve (AUC), and Matthew's Correlation Coefficient (MCC). Following the same scheme, another MLP was trained in selecting features with the disease label as the target. A comparison between the brain age-based classification and this procedure was performed.
Results:
The models trained on 100 features were selected since they showed the best performance during validation. Validation results for the brain age task presented an MAE, r and R² of 5.15 years 0.85, 0.70, respectively using 100 features, while test results were 5.08 years, 0.87 and 0.73 (See figure 2C). Validation for the classification task using the 100 brain age signature showed the next results:metrics among the 0.6-0.7 level Prec = 0.70, recall = 0.67, AUC = 0.83, MCC = 0.51 (see figure 2D). While on the test group results were, Prec = 0.63, recall = 0.63, AUC = 0.78, MCC = 0.45. Furthermore, these results are similar, although slightly inferior, to a classification task based on features selected on the disease labels (Test results 100 features: Prec = 0.64, recall = 0.66, AUC = 0.81, MCC = 0.49).
Conclusions:
Features selected for brain age estimation effectively classify conditions such as schizophrenia and chronic migraine. These brain age signatures accurately predict brain age and capture disease degeneration patterns. Future work could apply these methods in broader clinical settings, enhancing understanding of neurodegenerative diseases.
Lifespan Development:
Aging 1
Modeling and Analysis Methods:
Classification and Predictive Modeling 2
Novel Imaging Acquisition Methods:
Anatomical MRI
Keywords:
Computational Neuroscience
Modeling
Schizophrenia
1|2Indicates the priority used for review
Provide references using author date format
[2] Beheshti, I. (2020) "T1-weighted MRI-driven brain age estimation in Alzheimer’s disease and Parkinson’s disease." Aging and disease 11.3 618
[3] Cole, J. H. (2020), "Longitudinal assessment of multiple sclerosis with the brain‐age paradigm." Annals of neurology 88.1 93-105.
[4] Henschel, L. (2020), "Fastsurfer-a fast and accurate deep learning based neuroimaging pipeline." NeuroImage 219 117012.
[5] Klein, A. (2012), "101 labeled brain images and a consistent human cortical labeling protocol." Frontiers in neuroscience 6 171.
[6] Maia, F. (2022), "Effective sample size, dimensionality, and generalization in covariate shift adaptation." Neural Computing and Applications 1-13.
Acknowledgements
This work was supported by Ministerio de Ciencia e Innovación of Spain with research grants PID2021-124407NB-I00, TED2021-130758B-I00 and PRE2019-089176.