Adolescent Brain Cognitive Development Neurocognitive Prediction

The ABCD Neurocognitive Prediction Challenge is a community driven competition asking competitors to develop algorithms to predict fluid intelligence score from T1-w MRIs. In this work, we propose a deep learning combined with gradient boosting machine framework to solve this task. We train a convolutional neural network to compress the high dimensional MRI data and learn meaningful image features by predicting the 123 continuous-valued derived data provided with each MRI. These extracted features are then used to train a gradient boosting machine that predicts the residualized fluid intelligence score. Our approach achieved mean square error (MSE) scores of 18.4374, 68.7868, and 96.1806 for the training, validation, and test set respectively.

In this work, we utilize T1-weighted MR images and StackNet to predict fluid intelligence in adolescents. Our framework includes feature extraction, feature normalization, feature denoising, feature selection, training a StackNet, and predicting fluid intelligence. The extracted feature is the distribution of different brain tissues in different brain parcellation regions. The proposed StackNet consists of three layers and 11 models. Each layer uses the predictions from all previous layers including the input layer. The proposed StackNet is tested on a public benchmark Adolescent Brain Cognitive Development Neurocognitive Prediction Challenge 2019 and achieves a mean squared error of 82.42 on the combined training and validation set with 10-fold cross-validation. The proposed StackNet achieves a mean squared error of 94.25 on the testing data. The source code is available on GitHub (https://​github.​com/​UCSB-VRL/​ABCD-MICCAI2019).

Predicting fluid intelligence based on T1-weighted magnetic resonance imaging (MRI) scans poses several challenges, including developing an adequate data representation of three dimensional voxel data, extracting predictive information from this data representation, and devising a model that is able to leverage the predictive information. We evaluate two strategies for prediction of fluid intelligence given structural MRI scans acquired through the Adolescent Brain Cognitive Development (ABCD) Study: deep learning models trained on raw imagery and classical machine learning models trained on extracted features. Our best-performing solution consists of a classical machine learning model trained on a combination of provided brain volume estimates and extracted features. Specifically, a Gradient Boosting Regressor (GBR) trained on a PCA-reduced feature space produced the best performance (train MSE = 66.29, validation MSE = 70.16), surpassing regression models trained on the provided volume data alone, and 2D/3D Convolutional Neural Networks trained on various representations of imagery data. Nonetheless, these results remain slightly better than a mean prediction, suggesting that neither approach is capturing a high degree of variance in the data.

Brain morphometry derived from structural magnetic resonance imaging is a widely used quantitative biomarker in neuroimaging studies. In this paper, we investigate its usefulness for the Neurocognitive Prediction Challenge 2019.

An in-depth analysis of the features provided by the challenge (anatomical segmentation and volumes for regions of interest according to the SRI24 atlas) motivated us to process the native T1-weighted images with FreeSurfer 6.0, to derive reliable brain morphometry including surface based metrics. A combination of subcortical volumes and cortical thicknesses, curvatures, and surface areas was used as features for a support-vector regressor (SVR) to predict pre-residualized fluid intelligence scores. Results performing only slightly better than the baseline (uniformly predicting the mean) were observed on two internally held-out validation sets, while performance on the official validation set was approximately the same as the baseline.

Despite a large dataset of a specific cohort available for training, this suggests that structural brain morphometry alone has limited power for this challenge, at least with today’s imaging and post-processing methods.

We study predicting fluid intelligence of 9–10 year old children from T1-weighted magnetic resonance images. We extract volume and thickness measurements from MRI scans using FreeSurfer and the SRI24 atlas. We empirically compare two predictive models: (i) an ensemble of gradient boosted trees and (ii) a linear ridge regression model. For both, a Bayesian black-box optimizer for finding the best suitable prediction model is used. To systematically analyze feature importance our model, we employ results from game theory in the form of Shapley values. Our model with gradient boosting and FreeSurfer measures ranked third place among 24 submissions to the ABCD Neurocognitive Prediction Challenge. Our results on feature importance could be used to guide future research on the neurobiological mechanisms behind fluid intelligence in children.

The degree of association between fluid intelligence and neuroanatomy is important in refining our understanding of brain development. The primary goal of this work is twofold: to predict fluid intelligence from T1-weighed MRI, and to describe the MRI features that are associated with fluid intelligence. In this paper, we propose to ensemble the predictions of three machine learning strategies: Support Vector Machine (SVM), Random Forest (RF), and Bootstrapped Step Wise Model Selection (BSWiMS). Gender-stratified SVM was trained on children using age (ages 9–10), plus 122 volumetric scores provided by the ABCD challenge team. RF and BSWiMS were gender-stratified and trained using cubic root transformed data, summarized by left-right mean and relative absolute differences, and augmented by 19 volumetric statistical descriptors of major anatomical regions. Then, the transformed-augmented feature set was adjusted by age and the mean volume of the training set. The predictions of the three models were averaged to get the final prediction on each one of the test subjects. The Mean Squared Error (MSE) of MRI-predicted fluid intelligence on the test subjects was 100.89. The top features associated with fluid intelligence were the volume of the pons white mater and the volume of the parahippocampal gyrus.

We propose a four-layer fully-connected neural network (FNN) for predicting fluid intelligence scores from T1-weighted MR images for the ABCD-challenge. In addition to the volumes of brain structures, the FNN uses cortical WM/GM contrast and cortical thickness at 78 cortical regions. These last two measurements were derived from the T1-weighted MR images using cortical surfaces produced by the CIVET pipeline. The age and gender of the subjects and the scanner manufacturer are also used as features for the learning algorithm. This yielded 283 features provided to the FNN with two hidden layers of 20 and 15 nodes. The method was applied to the data from the ABCD study. Trained with a training set of 3736 subjects, the proposed method achieved a MSE of 71.596 and a correlation of 0.151 in the validation set of 415 subjects. For the final submission, the model was trained with 3568 subjects and it achieved a MSE of 94.0270 in the test set comprised of 4383 subjects.

Ensemble learning aggregates a set of models to solve the same problem and usually gives better results than a single model. We apply the ensemble method to seek a better prediction in the Adolescent Brain Cognitive Development Neurocognitive Prediction Challenge (ABCD-NP-Challenge). We manage to obtain a much better predicting accuracy on the fluid intelligence with the proposed ensemble method using volumetric data from T1w brain image than a single prediction model. In addition, we compare the results of adolescents with young adults using data from the Human Connectome Project (HCP). We find that raw fluid intelligence scores in HCP without regressing out covariates such as age and brain volume can be much better predicted by brain structure. Also, the prediction, in general, is more accurate in young adults than adolescents.

Decoding fluid intelligence from brain MRI data in adolescents is a highly challenging task. In this study, we took part in the ABCD Neurocognitive Prediction (NP) Challenge 2019, in which a large set of T1-weighted magnetic resonance imaging (MRI) data and pre-residualized fluid intelligence scores (corrected for brain volume, data collection site and sociodemographic variables) of children between 9–11 years were provided (\(N=3739\) for training, \(N=415\) for validation and \(N=4516\) for testing). We propose here the Caruana Ensemble Search method to choose best performing models over a large and diverse set of candidate models. These candidate models include convolutional neural networks (CNNs) applied to brain areas considered to be relevant in fluid intelligence (e.g. frontal and parietal areas) and high-performing standard machine learning methods (namely support vector regression, random forests, gradient boosting and XGBoost) applied to region-based scores including volume, mean intensity and count of gray matter voxels. To further create diversity and increase robustness, a wide set of hyperparameter configurations for each of the models was used. On the validation and the hold out test data, we obtained a mean squared error (MSE) of 71.15 and 93.68 respectively (rank 12 out of 24, MSE range 92.13–102.25). Among most selected models were XGBoost together with the three region-based scores, the other regression models together with volume or CNNs based on the middle frontal gyrus. We discuss these results in light of previous research findings on fluid intelligence.

Fluid intelligence (FI) indicates a set of general abilities like pattern recognition, abstract thinking, and problem-solving. FI is related to inherent, biological factors. We present a method to predict the fluid intelligence score in children (9–10 y/o) from their structural brain scans. For the purposes of this work, we used features derived from the T1-weighted Magnetic Resonance scans from the ABCD study. We used data from 3739 subjects for training and 415 for validation of the model. As features we used the volumes of gray matter regions of interest provided by the challenge organizers, as well as three additional groups of features. These include signal intensity features based on the ROIs, as well as shape-based features derived from the anterior and posterior cross sectional area of the corpus callosum. We used the random forest regressor model for prediction. We compare its performance to other regression-based models (XGBoost Regression and Support Vector Regression). Additionally, we ran a mean decrease accuracy (MDA) algorithm to select features that had high influence on the prediction results. The results we have obtained for the validation set were as follows: MSE = 67.39, R-squared = 0.0762. The proposed method showed promising results and has the potential to provide a good prediction of fluid intelligence based on structural brain scans.

In response to the ABCD Neurocognitive Prediction Challenge (ABCD-NP-Challenge 2019), we developed machine learning algorithms to predict the fluid intelligence (FI) score using T1-weighed magnetic resonance imaging (MRI) data. 122 volumetric scores of regions of interest from 3739 samples provided in the training set were used to train the models and 415 samples were assigned as validation samples. We performed feature reduction using principal factors factor analysis on the training set volume. 36 factors explaining 100% of the total variance were extracted; the top 18 explained 80% of the variance. We estimated three types of models based on: (1) all regional brain volumes, (2) the 18 top factors or (3) the 36 complete factors. We used Scikit-Learn’s grid search method to search the hyperparameter spaces of eight different machine learning algorithms. The best model, a Nu support vector regression model (NuSVR) using 36 factor scores as features, yielded the highest validation score (R² = 0.048) and a relatively low training score (0.22), the latter of which was important for reducing the degree of over-fitting. The mean squared errors (MSEs) for the training and validation samples were 68.2 and 68.6, respectively; the correlation coefficients were 0.54 and 0.21 (p < 0.0001 for both). The final MSE for the test set was 95.63. Learning curve analysis suggests that the current training sample size is still too small; increasing sample size should improve predictive accuracy. Overall, our results suggest that, given a large enough sample, machine learning methods with structural MRI data may be able to accurately estimate fluid intelligence.

We propose an AutoML approach for the prediction of fluid intelligence from T1-weighted magnetic resonance images. We extracted 122 features from MRI scans and employed Sequential Model-based Algorithm Configuration to search for the best prediction pipeline, including the best data pre-processing and regression model. In total, we evaluated over 2600 prediction pipelines. We studied our final model by employing results from game theory in the form of Shapley values. Results indicate that predicting fluid intelligence from volume measurements is a challenging task with many challenges. We found that our final ensemble of 50 prediction pipelines associated larger parahippocampal gyrus volumes with lower fluid intelligence, and higher pons white matter volume with higher fluid intelligence.

In this paper, we develop a 3D convolutional neural network to predict the fluid intelligence from T1-weighted MRI images by adding an encoder-decoder regularization. Considering that cerebellar volume is often highly correlated to intelligence of an individual, we propose to incorporate this morphological information into the framework for fluid intelligence prediction by utilizing an encoder-decoder regularization for brain structure segmentation simultaneously. Specifically, we first train an encoder-decoder network to generate the brain segmentation mask, where the discriminative morphological feature of the brain volume can be learned. Then, we reuse the encoder path of the network as the prediction network backbone for final fluid intelligence prediction by adding an additional regression part to predict the fluid intelligence value. The proposed framework is able to learn the discriminative relationship between the morphological information of brain structures and the intelligence score for more accurate prediction.

We predicted fluid intelligence from T1-weighted MRI data available as part of the ABCD NP Challenge 2019, using morphological similarity of grey-matter regions across the cortex. Individual structural covariance networks (SCN) were abstracted into graph-theory metrics averaged over nodes across the brain and in data-driven communities/modules. Metrics included degree, path length, clustering coefficient, centrality, rich club coefficient, and small-worldness. These features derived from the training set were used to build various regression models for predicting residual fluid intelligence scores, with performance evaluated both using cross-validation within the training set and using the held-out validation set. Our predictions on the test set were generated with a support vector regression model trained on the training set. We found minimal improvement over predicting a zero residual fluid intelligence score across the sample population, implying that structural covariance networks calculated from T1-weighted MR imaging data provide little information about residual fluid intelligence.

Prediction of cognitive performance from brain structural imaging data is a challenging machine learning topic. Participating in the ABCD Neurocognitive prediction challenge (2019), we implemented several machine learning models to solve this problem. Our results show superior performance from models relying on boosted decision trees and we find benefit from using two different sets of derived brain volumetric features. Lastly, across all models, we report an increase in performance by ensembling several different model types together in a final layer.

We applied several regression and deep learning methods to predict fluid intelligence scores from T1-weighted MRI scans as part of the ABCD Neurocognitive Prediction Challenge 2019. We used voxel intensities and probabilistic tissue-type labels derived from these as features to train the models. The best predictive performance (lowest mean-squared error) came from kernel ridge regression (\(\lambda =10\)), which produced a mean-squared error of 69.7204 on the validation set and 92.1298 on the test set. This placed our group in the fifth position on the validation leader board and first place on the final (test) leader board.

The ABCD Neurocognitive Prediction Challenge (ABCD-NP-Challenge 2019) made available T1-weighted structural scans for children alongside their fluid intelligence scores. The goal of the challenge was to use this anatomical brain data to train a model that could be successful in predicting fluid intelligence scores from held-out T1-weighted structural scans taken of other children. Functional magnetic resonance imaging (fMRI) has been moderately successful at identifying neural correlates of cognitive functioning, including intelligence. This study sought to leverage anatomical metrics within functionally defined regions, convolutional neural networks, and regression models to predict fluid intelligence. The proposed model performed competitively on the ABCD-NP-Challenge, and significantly outperformed a non deep-learning approach for behavior prediction based on the LASSO.

Fluid intelligence (Gf) has been defined as the ability to reason and solve previously unseen problems. Links to Gf have been found in magnetic resonance imaging (MRI) sequences such as functional MRI and diffusion tensor imaging. As part of the Adolescent Brain Cognitive Development Neurocognitive Prediction Challenge 2019, we sought to predict Gf in children aged 9–10 from T1-weighted (T1 W) MRIs. The data included atlas–aligned volumetric T1 W images, atlas–defined segmented regions, age, and sex for 3739 subjects used for training and internal validation and 415 subjects used for external validation. We trained sex-specific convolutional neural net (CNN) and random forest models to predict Gf. For the convolutional model, skull-stripped volumetric T1 W images aligned to the SRI24 brain atlas were used for training. Volumes of segmented atlas regions along with each subject’s age were used to train the random forest regressor models. Performance was measured using the mean squared error (MSE) of the predictions. Random forest models achieved lower MSEs than CNNs. Further, the external validation data had a better MSE for females than males (60.68 vs. 80.74), with a combined MSE of 70.83. Our results suggest that predictive models of Gf from volumetric T1 W MRI features alone may perform better when trained separately on male and female data. However, the performance of our models indicates that more information is necessary beyond the available data to make accurate predictions of Gf.

In this work, we aimed at predicting children’s fluid intelligence scores based on structural T1-weighted MR images from the largest long-term study of brain development and child health. The target variable was regressed on a data collection site, sociodemographic variables, and brain volume, thus being independent to the potentially informative factors, which were not directly related to the brain functioning. We investigated both feature extraction and deep learning approaches as well as different deep CNN architectures and their ensembles. We proposed an advanced architecture of VoxCNNs ensemble, which yields MSE (92.838) on a blind test.

From the ABCD dataset, we discover that besides the gray matter volume of cortical regions, other measures such as the mean cortical curvature, white matter volume and subcortical volume exhibit additional capabilities in the prediction of the pre-residulized fluid intelligence scores for adolescents. The MSE and R-square on validation dataset are improved from 70.65 and 0.0175 to 69.39 and 0.0350, respectively, comparing with using mostly the grey matter volume provided by the challenge organizer. Specifically, by employing a BlockPC-XGBoost framework we discover the following predictors in reducing the MSE on validation set: the gray matter volume of right posterior cingulate gyrus and left caudate nucleus, the entorhinal white matter volume of the left hemisphere, the number of detected surface holes, the globus pallidus volume, the mean curvatures of precentral gyrus, postcentral gyrus and Banks of Superior Temporal Sulcus.

We present a novel framework using 3D convolutional neural networks to predict residualized fluid intelligence scores in the MICCAI 2019 Adolescent Brain Cognitive Development Neurocognitive Prediction Challenge datasets. Using gray matter segmentations from T1-weighted MRI volumes as inputs, our framework identified several cortical and subcortical brain regions where the predicted errors were lower than random guessing in the validation set (mean squared error = 71.5252), and our final outcomes (mean squared error = 70.5787 in the validation set, 92.7407 in the test set) were comprised of the median scores predicted from these regions.