Background

The objective of this study was to investigate the potential of unsegmented structural T1w MR images of adolescent brain for predicting uncorrected/actual fluid intelligence scores without any predefined feature extraction. We also examined whether prediction of uncorrected scores is simply a harder problem from both biological and technical point of view, than prediction of residualised scores.

Methods

ABCD (Adolescent Brain Cognitive Development) study data was used from 7709 children aged 9–10, including T1-weighted MRIs and fluid intelligence scores, with data split into training (n = 3739), validation (n = 415) and test (n = 3555) subsets. We developed several deep learning convolutional neural network (CNN) models for both actual and residualised fluid intelligence score prediction from the MR images. State of the art, conventional or reverse 2D/3D CNN architectures were developed to perform the regression task and optimised based on Pearson's correlation coefficient, r. The models were then compared with published results on the same dataset.

Results

Our proposed model achieved prediction accuracies of r = 0.18 (p < 0.001) for the validation and r = 0.1 (p < 0.05) for the test, for actual IQ prediction. Our results showed that, although we achieved ~10 times higher correlation for the residualised score prediction than the correlations reported by previous CNN studies, using the same unsegmented MR images, it could not exceed the actual IQ prediction performance. This suggests that the image features associated with covariates aided up in the uncorrected score prediction rather than making the task harder.

Conclusion

Our deep learning CNN was able to establish a weak but stable correlation between structural brain features and raw fluid intelligence. To improve neuroimaging-based fluid intelligence prediction performance, future studies will be required to explore ensembled regression strategies with multiple machine learning algorithms on multimodal MRIs.

中文翻译：

---

用深度学习方法从结构 MRI 预测青春期的流体智力

背景

本研究的目的是调查未分割的青少年大脑结构 T1w MR 图像在没有任何预定义特征提取的情况下预测未校正/实际流体智力评分的潜力。我们还检查了从生物学和技术的角度来看，预测未修正分数是否比预测残差分数更难。

方法

ABCD（青少年大脑认知发展）研究数据来自 7709 名 9-10 岁儿童，包括 T1 加权 MRI 和流体智力评分，数据分为训练 ( n  = 3739)、验证 ( n  = 415) 和测试 ( n  = 3555) 子集。我们开发了几种深度学习卷积神经网络 (CNN) 模型，用于从 MR 图像中预测实际和残余流体智力分数。最先进的传统或反向 2D/3D CNN 架构被开发来执行回归任务并基于 Pearson 相关系数r进行优化。然后将模型与相同数据集上的已发布结果进行比较。

结果

我们提出的模型 在验证中实现了r  = 0.18 ( p < 0.001) 的预测精度， 在测试中实现了r  = 0.1 ( p < 0.05)，用于实际智商预测。我们的结果表明，虽然我们的残差分数预测的相关性比之前 CNN 研究报告的相关性高约 10 倍，但使用相同的未分割 MR 图像，它无法超过实际的 IQ 预测性能。这表明与协变量相关的图像特征有助于未校正分数预测，而不是使任务更难。

结论

我们的深度学习 CNN 能够在大脑结构特征和原始流体智力之间建立微弱但稳定的相关性。为了提高基于神经影像的流体智能预测性能，未来的研究将需要在多模态 MRI 上探索具有多种机器学习算法的集成回归策略。