Search

Understanding the genetic architecture of brain structure is challenging, partly due to difficulties in designing robust, non-biased descriptors of brain morphology. Until recently, brain measures for genome-wide association studies (GWAS) consisted of traditionally expert-defined or software-derived image-derived phenotypes (IDPs) that are often based on theoretical preconceptions or computed from limited amounts of data. Here, we present an approach to derive brain imaging phenotypes using unsupervised deep representation learning. We train a 3-D convolutional autoencoder model with reconstruction loss on 6130 UK Biobank (UKBB) participants' T1 or T2-FLAIR (T2) brain MRIs to create a 128-dimensional representation known as Unsupervised Deep learning derived Imaging Phenotypes (UDIPs). GWAS of these UDIPs in held-out UKBB subjects (n = 22,880 discovery and n = 12,359/11,265 replication cohorts for T1/T2) identified 9457 significant SNPs organized into 97 independent genetic loci of which 60 loci were replicated. Twenty-six loci were not reported in earlier T1 and T2 IDP-based UK Biobank GWAS. We developed a perturbation-based decoder interpretation approach to show that these loci are associated with UDIPs mapped to multiple relevant brain regions. Our results established unsupervised deep learning can derive robust, unbiased, heritable, and interpretable brain imaging phenotypes. A study utilizing unsupervised deep learning to generate interpretable brain imaging phenotypes from brain T1 and T2-FLAIR MRI identified 97 genetic loci enhancing understanding of brain structure genetics. [ABSTRACT FROM AUTHOR]

Deep learning has recently achieved best-in-class performance in several fields, including biomedical domains such as X-ray images. Yet, data scarcity poses a strict limit on training successful deep learning systems in many, if not most, biomedical applications, including those involving brain images. In this study, we translate state-of-the-art transfer learning techniques for single-subject prediction of simpler (sex and age) and more complex phenotypes (number of people in household, household income, fluid intelligence and smoking behavior). We fine-tuned 2D and 3D ResNet-18 convolutional neural networks for target phenotype predictions from brain images of ∼ 40,000 UK Biobank participants, after pretraining on YouTube videos from the Kinetics dataset and natural images from the ImageNet dataset. Transfer learning was effective on several phenotypes, especially sex and age classification. Additionally, transfer learning in particular outperformed deep learning models trained from scratch especially on smaller sample sizes. The out-of-sample performance using transfer learning from previously learned knowledge based on real-world images and videos could unlock the potential in many areas of imaging neuroscience where deep learning solutions are currently infeasible. • Neuroimaging faces challenges in effectively training DNNs due to data scarcity. • Transfer learning (TL) could aid in training DNNs in the data-scarce neuroimaging setting. • Using TL, we improved predictions on several phenotypes and smaller sample sizes. [ABSTRACT FROM AUTHOR]

• Individuals remain poorly predictable from population-level analyses of the brain. • The fidelity limits data scale, compute, and model flexibility impose are unknown. • We quantify these limits in multimodal, multitarget analyses across 23810 people. • We show a radical change in modelling regimes is needed for individual prediction. Our knowledge of the organisation of the human brain at the population-level is yet to translate into power to predict functional differences at the individual-level, limiting clinical applications and casting doubt on the generalisability of inferred mechanisms. It remains unknown whether the difficulty arises from the absence of individuating biological patterns within the brain, or from limited power to access them with the models and compute at our disposal. Here we comprehensively investigate the resolvability of such patterns with data and compute at unprecedented scale. Across 23 810 unique participants from UK Biobank, we systematically evaluate the predictability of 25 individual biological characteristics, from all available combinations of structural and functional neuroimaging data. Over 4526 GPU*hours of computation, we train, optimize, and evaluate out-of-sample 700 individual predictive models, including fully-connected feed-forward neural networks of demographic, psychological, serological, chronic disease, and functional connectivity characteristics, and both uni- and multi-modal 3D convolutional neural network models of macro- and micro-structural brain imaging. We find a marked discrepancy between the high predictability of sex (balanced accuracy 99.7%), age (mean absolute error 2.048 years, R2 0.859), and weight (mean absolute error 2.609Kg, R2 0.625), for which we set new state-of-the-art performance, and the surprisingly low predictability of other characteristics. Neither structural nor functional imaging predicted an individual's psychology better than the coincidence of common chronic disease (p < 0.05). Serology predicted chronic disease (p < 0.05) and was best predicted by it (p < 0.001), followed by structural neuroimaging (p < 0.05). Our findings suggest either more informative imaging or more powerful models will be needed to decipher individual level characteristics from the human brain. We make our models and code openly available. [ABSTRACT FROM AUTHOR]

Convolutional neural networks (CNNs)—as a type of deep learning—have been specifically designed for highly heterogeneous data, such as natural images. Neuroimaging data, however, is comparably homogeneous due to (1) the uniform structure of the brain and (2) additional efforts to spatially normalize the data to a standard template using linear and non-linear transformations. To harness spatial homogeneity of neuroimaging data, we suggest here a new CNN architecture that combines the idea of hierarchical abstraction in CNNs with a prior on the spatial homogeneity of neuroimaging data. Whereas early layers are trained globally using standard convolutional layers, we introduce patch individual filters (PIF) for higher, more abstract layers. By learning filters in individual latent space patches without sharing weights, PIF layers can learn abstract features faster and specific to regions. We thoroughly evaluated PIF layers for three different tasks and data sets, namely sex classification on UK Biobank data, Alzheimer's disease detection on ADNI data and multiple sclerosis detection on private hospital data, and compared it with two baseline models, a standard CNN and a patch-based CNN. We obtained two main results: First, CNNs using PIF layers converge consistently faster, measured in run time in seconds and number of iterations than both baseline models. Second, both the standard CNN and the PIF model outperformed the patch-based CNN in terms of balanced accuracy and receiver operating characteristic area under the curve (ROC AUC) with a maximal balanced accuracy (ROC AUC) of 94.21% (99.10%) for the sex classification task (PIF model), and 81.24% and 80.48% (88.89% and 87.35%) respectively for the Alzheimer's disease and multiple sclerosis detection tasks (standard CNN model). In conclusion, we demonstrated that CNNs using PIF layers result in faster convergence while obtaining the same predictive performance as a standard CNN. To the best of our knowledge, this is the first study that introduces a prior in form of an inductive bias to harness spatial homogeneity of neuroimaging data. [ABSTRACT FROM AUTHOR]