MEGnet: Automatic ICA-based artifact removal for MEG using spatiotemporal convolutional neural networks.

Magnetoencephalography (MEG) is a functional neuroimaging tool that records the magnetic fields induced by neuronal activity; however, signal from non-neuronal sources can corrupt the data. Eye-blinks, saccades, and cardiac activity are three of the most common sources of non-neuronal artifacts. They can be measured by affixing eye proximal electrodes, as in electrooculography (EOG), and chest electrodes, as in electrocardiography (ECG), however this complicates imaging setup, decreases patient comfort, and can induce further artifacts from movement. This work proposes an EOG- and ECG-free approach to identify eye-blinks, saccades, and cardiac activity signals for automated artifact suppression. The contribution of this work is three-fold. First, using a data driven, multivariate decomposition approach based on Independent Component Analysis (ICA), a highly accurate artifact classifier is constructed as an amalgam of deep 1-D and 2-D Convolutional Neural Networks (CNNs) to automate the identification and removal of ubiquitous whole brain artifacts including eye-blink, saccade, and cardiac artifacts. The specific architecture of this network is optimized through an unbiased, computer-based hyperparameter random search. Second, visualization methods are applied to the learned abstraction to reveal what features the model uses and to bolster user confidence in the model's training and potential for generalization. Finally, the model is trained and tested on both resting-state and task MEG data from 217 subjects, and achieves a new state-of-the-art in artifact detection accuracy of 98.95% including 96.74% sensitivity and 99.34% specificity on the held out test-set. This work automates MEG processing for both clinical and research use, adapts to the acquired acquisition time, and can obviate the need for EOG or ECG electrodes for artifact detection.

Longitudinal prognosis of Parkinson's outcomes using causal connectivity.

Despite the prevalence of Parkinson's disease (PD), there are no clinically-accepted neuroimaging biomarkers to predict the trajectory of motor or cognitive decline or differentiate Parkinson's disease from atypical progressive parkinsonian diseases. Since abnormal connectivity in the motor circuit and basal ganglia have been previously shown as early markers of neurodegeneration, we hypothesize that patterns of interregional connectivity could be useful to form patient-specific predictive models of disease state and of PD progression. We use fMRI data from subjects with Multiple System Atrophy (MSA), Progressive Supranuclear Palsy (PSP), idiopathic PD, and healthy controls to construct predictive models for motor and cognitive decline and differentiate between the four subgroups. Further, we identify the specific connections most informative for progression and diagnosis. When predicting the one-year progression in the MDS-UPDRS-III1* and Montreal Cognitive assessment (MoCA), we achieve new state-of-the-art mean absolute error performance. Additionally, the balanced accuracy we achieve in the diagnosis of PD, MSA, PSP, versus healthy controls surpasses that attained in most clinics, underscoring the relevance of the brain connectivity features. Our models reveal the connectivity between deep nuclei, motor regions, and the thalamus as the most important for prediction. Collectively these results demonstrate the potential of fMRI connectivity as a prognostic biomarker for PD and increase our understanding of this disease.

Adversarially-Regularized Mixed Effects Deep Learning (ARMED) Models Improve Interpretability, Performance, and Generalization on Clustered (non-iid) Data.

Natural science datasets frequently violate assumptions of independence. Samples may be clustered (e.g., by study site, subject, or experimental batch), leading to spurious associations, poor model fitting, and confounded analyses. While largely unaddressed in deep learning, this problem has been handled in the statistics community through mixed effects models, which separate cluster-invariant fixed effects from cluster-specific random effects. We propose a general-purpose framework for Adversarially-Regularized Mixed Effects Deep learning (ARMED) models through non-intrusive additions to existing neural networks: 1) an adversarial classifier constraining the original model to learn only cluster-invariant features, 2) a random effects subnetwork capturing cluster-specific features, and 3) an approach to apply random effects to clusters unseen during training. We apply ARMED to dense, convolutional, and autoencoder neural networks on 4 datasets including simulated nonlinear data, dementia prognosis and diagnosis, and live-cell image analysis. Compared to prior techniques, ARMED models better distinguish confounded from true associations in simulations and learn more biologically plausible features in clinical applications. They can also quantify inter-cluster variance and visualize cluster effects in data. Finally, ARMED matches or improves performance on data from clusters seen during training (5-28% relative improvement) and generalization to unseen clusters (2-9% relative improvement) versus conventional models.

Reproducible neuroimaging features for diagnosis of autism spectrum disorder with machine learning.

Autism spectrum disorder (ASD) is the fourth most common neurodevelopmental disorder, with a prevalence of 1 in 160 children. Accurate diagnosis relies on experts, but such individuals are scarce. This has led to increasing interest in the development of machine learning (ML) models that can integrate neuroimaging features from functional and structural MRI (fMRI and sMRI) to help reveal central nervous system alterations characteristic of ASD. We optimized and compared the performance of 12 of the most popular and powerful ML models. Each was separately trained using 15 different combinations of fMRI and sMRI features and optimized with an unbiased model search. Deep learning models predicted ASD with the highest diagnostic accuracy and generalized well to other MRI datasets. Our model achieves state-of-the-art 80% area under the ROC curve (AUROC) in diagnosis on test data from the IMPAC dataset; and 86% and 79% AUROC on the external ABIDE I and ABIDE II datasets (with further improvement to 93% and 90% after supervised domain adaptation). The highest performing models identified reproducible putative biomarkers for accurate ASD diagnosis in accord with known ASD markers as well as novel cerebellar biomarkers. Such reproducibility lends credence to their tremendous potential for defining and using a set of truly generalizable ASD biomarkers that will advance scientific understanding of neuronal changes in ASD.

Association between resting-state functional brain connectivity and gene expression is altered in autism spectrum disorder.

Gene expression covaries with brain activity as measured by resting state functional magnetic resonance imaging (MRI). However, it is unclear how genomic differences driven by disease state can affect this relationship. Here, we integrate from the ABIDE I and II imaging cohorts with datasets of gene expression in brains of neurotypical individuals and individuals with autism spectrum disorder (ASD) with regionally matched brain activity measurements from fMRI datasets. We identify genes linked with brain activity whose association is disrupted in ASD. We identified a subset of genes that showed a differential developmental trajectory in individuals with ASD compared with controls. These genes are enriched in voltage-gated ion channels and inhibitory neurons, pointing to excitation-inhibition imbalance in ASD. We further assessed differences at the regional level showing that the primary visual cortex is the most affected region in ASD. Our results link disrupted brain expression patterns of individuals with ASD to brain activity and show developmental, cell type, and regional enrichment of activity linked genes.

Patterns of Pretreatment Reward Task Brain Activation Predict Individual Antidepressant Response: Key Results From the EMBARC Randomized Clinical Trial.

BACKGROUND: The lack of biomarkers to inform antidepressant selection is a key challenge in personalized depression treatment. This work identifies candidate biomarkers by building deep learning predictors of individual treatment outcomes using reward processing measures from functional magnetic resonance imaging, clinical assessments, and demographics. METHODS: Participants in the EMBARC (Establishing Moderators and Biosignatures of Antidepressant Response in Clinical Care) study (n = 222) underwent reward processing task-based functional magnetic resonance imaging at baseline and were randomized to 8 weeks of sertraline (n = 106) or placebo (n = 116). Subsequently, sertraline nonresponders (n = 37) switched to 8 weeks of bupropion. The change in Hamilton Depression Rating Scale was measured after treatment. Reward processing, clinical measurements, and demographics were used to train treatment-specific deep learning models. RESULTS: The predictive model for sertraline achieved R2 of 48% (95% CI, 33%-61%; p < 10-3) in predicting the change in Hamilton Depression Rating Scale and number-needed-to-treat (NNT) of 4.86 participants in predicting response. The placebo model achieved R2 of 28% (95% CI, 15%-42%; p < 10-3) and NNT of 2.95 in predicting response. The bupropion model achieved R2 of 34% (95% CI, 10%-59%, p < 10-3) and NNT of 1.68 in predicting response. Brain regions where reward processing activity was predictive included the prefrontal cortex and cerebellar crus 1 for sertraline and the cingulate cortex, caudate, orbitofrontal cortex, and crus 1 for bupropion. CONCLUSIONS: These findings demonstrate the utility of reward processing measurements and deep learning to predict antidepressant outcomes and to form multimodal treatment biomarkers.

Glaucoma Diagnosis Through the Integration of Optical Coherence Tomography/Angiography and Machine Learning Diagnostic Models.

Purpose: To establish optical coherence tomography (OCT)/angiography (OCTA) parameter ranges for healthy eyes (HE) and glaucomatous eyes (GE) for a North Texas based population; to develop a machine learning (ML) tool and to identify the most accurate diagnostic parameters for clinical glaucoma diagnosis. Patients and Methods: In this retrospective cross-sectional study, we included 1371 eligible eyes, 462 HE and 909 GE (377 ocular hypertension, 160 mild, 156 moderate, 216 severe), from 735 subjects. Demographic data and full OCTA parameters were collected. A Kruskal-Wallis test was used to produce the normative database. Models were trained to solve a two-class problem (HE vs GE) and four-class problem (HE vs mild vs moderate vs severe GE). A rigorous nested, stratified, group, 5×10 fold cross-validation strategy was applied to partition the data. Six ML algorithms were compared using classical and deep learning approaches. Over 2500 ML models were optimized using random search, with performance compared using mean validation accuracy. Final performance was reported on held-out test data using accuracy and F1 score. Decision trees and feature importance were produced for the final model. Results: We found differences across glaucoma severities for age, gender, hypertension, Black and Asian race, and all OCTA parameters, except foveal avascular zone area and perimeter (p<0.05). The XGBoost algorithm achieved the highest test performance for both the two-class (F1 score 83.8%; accuracy 83.9%; standard deviation 0.03%) and four-class (F1 score 62.4%; accuracy 71.3%; standard deviation 0.013%) problem. A set of interpretable decision trees provided the most important predictors of the final model; inferior temporal and inferior hemisphere vessel density and peripapillary retinal nerve fiber layer thickness were identified as key diagnostic parameters. Conclusion: This study established a normative database for our North Texas based population and created ML tools utilizing OCT/A that may aid clinicians in glaucoma management.

Predicting Parkinson's disease trajectory using clinical and neuroimaging baseline measures.

INTRODUCTION: Predictive biomarkers of Parkinson's Disease progression are needed to expedite neuroprotective treatment development and facilitate prognoses for patients. This work uses measures derived from resting-state functional magnetic resonance imaging, including regional homogeneity (ReHo) and fractional amplitude of low frequency fluctuations (fALFF), to predict an individual's current and future severity over up to 4 years and to elucidate the most prognostic brain regions. METHODS: ReHo and fALFF are measured for 82 Parkinson's Disease subjects and used to train machine learning predictors of baseline clinical and future severity at 1 year, 2 years, and 4 years follow-up as measured by the Movement Disorder Society Unified Parkinson's Disease Rating Scale (MDS-UPDRS). Predictive performance is measured with nested cross-validation, validated on an external dataset, and again validated through leave-one-site-out cross-validation. Important predictive features are identified. RESULTS: The models explain up to 30.4% of the variance in current MDS-UPDRS scores, 55.8% of the variance in year 1 scores, and 47.1% of the variance in year 2 scores (p < 0.0001). For distinguishing high and low-severity individuals at each timepoint (MDS-UPDRS score above or below the median, respectively), the models achieve positive predictive values up to 79% and negative predictive values up to 80%. Higher ReHo and fALFF in several regions, including components of the default motor network, predicted lower severity across current and future timepoints. CONCLUSION: These results identify an accurate prognostic neuroimaging biomarker which may be used to better inform enrollment in trials of neuroprotective treatments and enable physicians to counsel their patients.

Architectural configurations, atlas granularity and functional connectivity with diagnostic value in Autism Spectrum Disorder.

Currently, the diagnosis of Autism Spectrum Disorder (ASD) is dependent upon a subjective, time-consuming evaluation of behavioral tests by an expert clinician. Non-invasive functional MRI (fMRI) characterizes brain connectivity and may be used to inform diagnoses and democratize medicine. However, successful construction of predictive models, such as deep learning models, from fMRI requires addressing key choices about the model's architecture, including the number of layers and number of neurons per layer. Meanwhile, deriving functional connectivity (FC) features from fMRI requires choosing an atlas with an appropriate level of granularity. Once an accurate diagnostic model has been built, it is vital to determine which features are predictive of ASD and if similar features are learned across atlas granularity levels. Identifying new important features extends our understanding of the biological underpinnings of ASD, while identifying features that corroborate past findings and extend across atlas levels instills model confidence. To identify aptly suited architectural configurations, probability distributions of the configurations of high versus low performing models are compared. To determine the effect of atlas granularity, connectivity features are derived from atlases with 3 levels of granularity and important features are ranked with permutation feature importance. Results show the highest performing models use between 2-4 hidden layers and 16-64 neurons per layer, granularity dependent. Connectivity features identified as important across all 3 atlas granularity levels include FC to the supplementary motor gyrus and language association cortex, regions whose abnormal development are associated with deficits in social and sensory processing common in ASD. Importantly, the cerebellum, often not included in functional analyses, is also identified as a region whose abnormal connectivity is highly predictive of ASD. Results of this study identify important regions to include in future studies of ASD, help assist in the selection of network architectures, and help identify appropriate levels of granularity to facilitate the development of accurate diagnostic models of ASD.

Anatomically-Informed Data Augmentation for Functional MRI with Applications to Deep Learning.

The application of deep learning to build accurate predictive models from functional neuroimaging data is often hindered by limited dataset sizes. Though data augmentation can help mitigate such training obstacles, most data augmentation methods have been developed for natural images as in computer vision tasks such as CIFAR, not for medical images. This work helps to fills in this gap by proposing a method for generating new functional Magnetic Resonance Images (fMRI) with realistic brain morphology. This method is tested on a challenging task of predicting antidepressant treatment response from pre-treatment task-based fMRI and demonstrates a 26% improvement in performance in predicting response using augmented images. This improvement compares favorably to state-of-the-art augmentation methods for natural images. Through an ablative test, augmentation is also shown to substantively improve performance when applied before hyperparameter optimization. These results suggest the optimal order of operations and support the role of data augmentation method for improving predictive performance in tasks using fMRI.

Multiple Deep Learning Architectures Achieve Superior Performance Diagnosing Autism Spectrum Disorder Using Features Previously Extracted from Structural and Functional MRI.

The diagnosis of Autism Spectrum Disorder (ASD) is a subjective process requiring clinical expertise in neurodevelopmental disorders. Since such expertise is not available at many clinics, automated diagnosis using machine learning (ML) algorithms would be of great value to both clinicians and the imaging community to increase the diagnoses' availability and reproducibility while reducing subjectivity. This research systematically compares the performance of classifiers using over 900 subjects from the IMPAC database [1], using the database's derived anatomical and functional features to diagnose a subject as autistic or healthy. In total 12 classifiers are compared from 3 categories including: 6 nonlinear shallow ML models, 3 linear shallow models, and 3 deep learning models. When evaluated with an AUC ROC performance metric, results include: (1) amongst the shallow learning methods, linear models outperformed nonlinear models, agreeing with [2]. (2) Deep learning models outperformed shallow ML models. (3) The best model was a dense feedforward network, achieving 0.80 AUC which compares to the recently reported 0.79±0.01 AUC average of the top 10 methods from the IMPAC challenge [3]. These results demonstrate that even when using features derived from imaging data, deep learning methods can provide additional predictive accuracy over classical methods.

Deep Learning Convolutional Neural Networks for the Estimation of Liver Fibrosis Severity from Ultrasound Texture.

Diagnosis and staging of liver fibrosis is a vital prognostic marker in chronic liver diseases. Due to the inaccuracies and risk of complications associated with liver core needle biopsy, the current standard for diagnosis, other less invasive methods are sought for diagnosis. One such method that has been shown to correlate well with liver fibrosis is shear wave velocity measured by ultrasound (US) shear wave elastography; however, this technique requires specific software, hardware, and training. A current perspective in the radiology community is that the texture pattern from an US image may be predictive of the stage of liver fibrosis. We propose the use of convolutional neural networks (CNNs), a framework shown to be well suited for real world image interpretation, to test whether the texture pattern in gray scale elastography images (B-mode US with fixed, subject-agnostic acquisition settings) is predictive of the shear wave velocity (SWV). In this study, gray scale elastography images from over 300 patients including 3,500 images with corresponding SWV measurements were preprocessed and used as input to 100 different CNN architectures that were trained to regress shear wave velocity. In this study, even the best performing CNN explained only negligible variation in the shear wave velocity measures. These extensive test results suggest that the gray scale elastography image texture provides little predictive information about shear wave velocity and liver fibrosis.

Predicting Response to the Antidepressant Bupropion using Pretreatment fMRI.

Major depressive disorder is a primary cause of disability in adults with a lifetime prevalence of 6-21% worldwide. While medical treatment may provide symptomatic relief, response to any given antidepressant is unpredictable and patient-specific. The standard of care requires a patient to sequentially test different antidepressants for 3 months each until an optimal treatment has been identified. For 30-40% of patients, no effective treatment is found after more than one year of this trial-and-error process, during which a patient may suffer loss of employment or marriage, undertreated symptoms, and suicidal ideation. This work develops a predictive model that may be used to expedite the treatment selection process by identifying for individual patients whether the patient will respond favorably to bupropion, a widely prescribed antidepressant, using only pretreatment imaging data. This is the first model to do so for individuals for bupropion. Specifically, a deep learning predictor is trained to estimate the 8-week change in Hamilton Rating Scale for Depression (HAMD) score from pretreatment task-based functional magnetic resonance imaging (fMRI) obtained in a randomized controlled antidepressant trial. An unbiased neural architecture search is conducted over 800 distinct model architecture and brain parcellation combinations, and patterns of model hyperparameters yielding the highest prediction accuracy are revealed. The winning model identifies bupropion-treated subjects who will experience remission with the number of subjects needed-to-treat (NNT) to lower morbidity of only 3.2 subjects. It attains a substantially high neuroimaging study effect size explaining 26% of the variance (R2 = 0.26) and the model predicts post-treatment change in the 52-point HAMD score with an RMSE of 4.71. These results support the continued development of fMRI and deep learning-based predictors of response for additional depression treatments.