Transfer learning approaches in classification of mental disorders using neuroimaging

Abstract

Resting-state functional magnetic resonance imaging (rs-fMRI) data is widely used to characterize brain function in health and disease. Specifically, brain networks obtained from rs-fMRI based functional connectivity (FC) have been used to investigate disconnection syndromes in psychiatry. A good example is Autism Spectrum Disorder (ASD), wherein statistical group differences between ASD and controls in terms of FC have been widely reported, with FC being predominantly weaker in ASD. The aberrant behaviors of ASD patients have been found to be associated with abnormal patterns of RSFC, including the reduced correlation between frontal and posterior brain networks and local FC strengthening with long-distance FC reduction. Statistical group comparison suffers from the disadvantage that it does not possess the ability to predict the outcome (such as diagnostic status) in a novel subject. Machine learning models have been used for individual subject-level characterization as an alternative. Deep learning models outperform traditional machine learning methods in diagnostic classification. For ASD, open-source datasets such as ABIDE (Autism Brain Imaging Data Exchange) have accelerated the application of deep learning to the diagnostic category. However, overfitting is the main issue that constrains the validity and generalizability of deep learning and traditional machine learning. Overfitting refers as a well-trained deep neural network that can achieve high prediction accuracy in the training dataset but has poor prediction accuracy in the unseen test dataset. The primary cause of overfitting is the relatively small sample size in training dataset compared to the high dimensionality of the feature space, known as the “curse of dimensionality”. In addition, to address the sample size issue, data is being increasingly aggregated across sites to form large databases such as ABIDE, given that acquiring a large number of subjects from a single site can be costly and time-consuming. However, the problem with such an approach is wide inter-site variability in data characteristics that are non-neural in origin: differences across MRI vendors, pulse sequences, sequence parameters, sampling of patient populations, and data preprocessing pipelines. Consequently, machine learning (ML) approaches, including deep learning (DL), tend to produce high accuracy in diagnostic classification using single-site data but poor accuracy using multi-sites data. This lack of generalizability in the model is the most significant barrier to adopting ML and DL in neuroimaging-based diagnostics. We propose two approaches that address this issue. We used a VAE-CNN (variational autoencoder convolutional neural network) transfer learning model in the first project. Because it is harder to acquire and aggregate patient population data than healthy controls, comparatively larger samples are available from healthy controls in the public domain, we propose to address overfitting by using larger healthy samples to learn the neural signature of healthy controls, with the aim of “transferring” that learning into the context of discriminating clinical populations. Here, we investigate the utility of transfer learning from HCP (human connectome project) healthy control data for improving the classification of individuals with ASD from their healthy peers in ABIDE data. We identify the biomarkers contributing to classification using the Layer-Wise Relevance Propagation (LRP) algorithm. Results show that the proposed transfer learning method outperforms state-of-the-art deep learning methods in ASD classification, especially when training and testing data are drawn from different data acquisition sites. In the second project, we propose domain adaptation for improving the generalizability of neuroimaging-based diagnostic classification. Domain adaptation aims to improve classification performance in a given target domain by utilizing the knowledge learned from a different source domain by making data distributions of the two domains as similar as possible. To validate the utility of domain adaptation for classifying multi-site fMRI data, we developed a variational autoencoder – maximum mean discrepancy (VAE-MMD) model for three-way diagnostic classification of Autism, Asperger’s syndrome, and controls. In domain adaptation, we chose ABIDEII (Autism Brain Imaging Data Exchange) as the target domain data and ABIDEI as the source domain data. The results show that the domain adaptation approach achieved superior test accuracy of ABIDEII compared to baseline methods using just ABIDEII for classification. In addition, we augmented the source domain with additional healthy control subjects from Healthy Brain Network (HBN) and Amsterdam Open MRI Collection (AOMIC) datasets, enabling transfer learning to improve classification performance. Finally, we compared domain adaptation and combined statistical ComBat harmonization in this study. The result demonstrated that the domain adaptation model could be improved when combined with statistical methods. We openly share our data and model so that the neuroimaging community can explore the possibility of further improvement of the model by utilizing the ever-increasing amount of healthy control fMRI data in the public domain.