Prediction of fluid intelligence from T1-w MRI images: A precise two-step deep learning framework

Abstract

The Adolescent Brain Cognitive Development (ABCD) Neurocognitive Prediction Challenge (ABCD-NP-Challenge) is a community-driven competition that challenges competitors to develop algorithms to predict fluid intelligence scores from T1-w MRI images. In this work, a two-step deep learning pipeline is proposed to improve the prediction accuracy of fluid intelligence scores. In terms of the first step, the main contributions of this study include the following: (1) the concepts of the residual network (ResNet) and the squeeze-and-excitation network (SENet) are utilized to improve the original 3D U-Net; (2) in the segmentation process, the pixels in symmetrical brain regions are assigned the same label; (3) to remove redundant background information from the segmented regions of interest (ROIs), a minimum bounding cube (MBC) is used to enclose the ROIs. This new segmentation structure can greatly improve the segmentation performance of the ROIs in the brain as compared with the classical convolutional neural network (CNN), which yields a Dice coefficient of 0.8920. In the second stage, MBCs are used to train neural network regression models for enhanced nonlinearity. The fluid intelligence score prediction results of the proposed method are found to be superior to those of current state-of-the-art approaches, and the proposed method achieves a mean square error (MSE) of 82.56 on a test data set, which reflects a very competitive performance.

Can deep learning predict human intelligence from structural brain MRI?

“…As we mentioned in section that several studies used sMRI-based regional brain volumes as features in different machine learning methods to predict intelligence scores 4,[6][7][8][9][10][11][12][13][14][15][16][17][18][19][20][21][22][27][28][29][30][31] . These studies used ∼8,500 healthy subjects for model training and then predicted the residual PIQ score of more than 3,500 adolescents with a mean square error (MSE) ranging from 86 to 103 (for a range of true residual PIQ score of [−40, 30]), or a correlation of 10% (p < 0.05).…”

Imaging biomarkers on Angio-CT for predicting the efficacy of transarterial chemoembolization in hepatocellular carcinoma

“…As an essential treatment for advanced HCC (4), it is vital to predicting the efficacy of TACE as early as possible. It is a key trigger in predicting therapeutic efficacy to consider whether the whole tumor is included in the TACE procedure or not (21). For HCC nodules without being wholly included in the TACE target, the efficacy of TACE will be incomplete, leading to early local recurrence (22).…”

Can deep learning predict human intelligence from structural brain MRI?

“…We also computed the residual scores to account for the influence of covariates (see below). Deep learning strategies that predict absolute and residual scores have been used in other machine learning studies 4,[6][7][8][9][10][11][12][13][14][15][16][17][18][19][20][21][22][35][36][37][38][39] .…”

“…where q res and q are residual and absolute FSIQ/PIQ/VIQ scores, respectively, A denotes age in years, S denotes sex (1: male, 2: female), D denotes the diagnostic group (1: healthy, 2: autistic), E : (1 ≤ E ≤ 15) denotes the sample collection sites, and α, β , γ, δ , and η are parameters of linear regression. In a recent challenge to predict residual fluid intelligence (related to PIQ) in ∼ 8,500 healthy adolescents of 9-10 years 4,[6][7][8][9][10][11][12][13][14][15][16][17][18][19][20][21][22][35][36][37][38][39] , sMRI data were residualized using brain volume, data collection site, age, sex, race/ethnicity, highest parental education, parental income and parental marital status as independent variables. Since the data set we used in this study (i.e., ABIDE I) does not contain information on race/ethnicity, highest parental education, parental income, and parental marital status, we only used the rest of the factors as independent variables in Eq.…”

“…Similar to other brain MRI-based sex, age, and neurocognition estimation works 4,[6][7][8][9][10][11][12][13][14][15][16][17][18][19][20][21][22][33][34][35][36][37][38][39] , we used the mean absolute error (MAE) loss function to train our CNNs defined as:…”