The ReCoDe addiction research consortium: Losing and regaining control over drug intake-Findings and future perspectives.

Substance use disorders (SUDs) are seen as a continuum ranging from goal-directed and hedonic drug use to loss of control over drug intake with aversive consequences for mental and physical health and social functioning. The main goals of our interdisciplinary German collaborative research centre on Losing and Regaining Control over Drug Intake (ReCoDe) are (i) to study triggers (drug cues, stressors, drug priming) and modifying factors (age, gender, physical activity, cognitive functions, childhood adversity, social factors, such as loneliness and social contact/interaction) that longitudinally modulate the trajectories of losing and regaining control over drug consumption under real-life conditions. (ii) To study underlying behavioural, cognitive and neurobiological mechanisms of disease trajectories and drug-related behaviours and (iii) to provide non-invasive mechanism-based interventions. These goals are achieved by: (A) using innovative mHealth (mobile health) tools to longitudinally monitor the effects of triggers and modifying factors on drug consumption patterns in real life in a cohort of 900 patients with alcohol use disorder. This approach will be complemented by animal models of addiction with 24/7 automated behavioural monitoring across an entire disease trajectory; i.e. from a naïve state to a drug-taking state to an addiction or resilience-like state. (B) The identification and, if applicable, computational modelling of key molecular, neurobiological and psychological mechanisms (e.g., reduced cognitive flexibility) mediating the effects of such triggers and modifying factors on disease trajectories. (C) Developing and testing non-invasive interventions (e.g., Just-In-Time-Adaptive-Interventions (JITAIs), various non-invasive brain stimulations (NIBS), individualized physical activity) that specifically target the underlying mechanisms for regaining control over drug intake. Here, we will report on the most important results of the first funding period and outline our future research strategy.

Deep neural network heatmaps capture Alzheimer's disease patterns reported in a large meta-analysis of neuroimaging studies.

Deep neural networks currently provide the most advanced and accurate machine learning models to distinguish between structural MRI scans of subjects with Alzheimer's disease and healthy controls. Unfortunately, the subtle brain alterations captured by these models are difficult to interpret because of the complexity of these multi-layer and non-linear models. Several heatmap methods have been proposed to address this issue and analyze the imaging patterns extracted from the deep neural networks, but no quantitative comparison between these methods has been carried out so far. In this work, we explore these questions by deriving heatmaps from Convolutional Neural Networks (CNN) trained using T1 MRI scans of the ADNI data set and by comparing these heatmaps with brain maps corresponding to Support Vector Machine (SVM) activation patterns. Three prominent heatmap methods are studied: Layer-wise Relevance Propagation (LRP), Integrated Gradients (IG), and Guided Grad-CAM (GGC). Contrary to prior studies where the quality of heatmaps was visually or qualitatively assessed, we obtained precise quantitative measures by computing overlap with a ground-truth map from a large meta-analysis that combined 77 voxel-based morphometry (VBM) studies independently from ADNI. Our results indicate that all three heatmap methods were able to capture brain regions covering the meta-analysis map and achieved better results than SVM activation patterns. Among them, IG produced the heatmaps with the best overlap with the independent meta-analysis.

Artificial Intelligence Bias in Health Care: Web-Based Survey.

BACKGROUND: Resources are increasingly spent on artificial intelligence (AI) solutions for medical applications aiming to improve diagnosis, treatment, and prevention of diseases. While the need for transparency and reduction of bias in data and algorithm development has been addressed in past studies, little is known about the knowledge and perception of bias among AI developers. OBJECTIVE: This study's objective was to survey AI specialists in health care to investigate developers' perceptions of bias in AI algorithms for health care applications and their awareness and use of preventative measures. METHODS: A web-based survey was provided in both German and English language, comprising a maximum of 41 questions using branching logic within the REDCap web application. Only the results of participants with experience in the field of medical AI applications and complete questionnaires were included for analysis. Demographic data, technical expertise, and perceptions of fairness, as well as knowledge of biases in AI, were analyzed, and variations among gender, age, and work environment were assessed. RESULTS: A total of 151 AI specialists completed the web-based survey. The median age was 30 (IQR 26-39) years, and 67% (101/151) of respondents were male. One-third rated their AI development projects as fair (47/151, 31%) or moderately fair (51/151, 34%), 12% (18/151) reported their AI to be barely fair, and 1% (2/151) not fair at all. One participant identifying as diverse rated AI developments as barely fair, and among the 2 undefined gender participants, AI developments were rated as barely fair or moderately fair, respectively. Reasons for biases selected by respondents were lack of fair data (90/132, 68%), guidelines or recommendations (65/132, 49%), or knowledge (60/132, 45%). Half of the respondents worked with image data (83/151, 55%) from 1 center only (76/151, 50%), and 35% (53/151) worked with national data exclusively. CONCLUSIONS: This study shows that the perception of biases in AI overall is moderately fair. Gender minorities did not once rate their AI development as fair or very fair. Therefore, further studies need to focus on minorities and women and their perceptions of AI. The results highlight the need to strengthen knowledge about bias in AI and provide guidelines on preventing biases in AI health care applications.

Stress-induced brain activity, brain atrophy, and clinical disability in multiple sclerosis.

Prospective clinical studies support a link between psychological stress and multiple sclerosis (MS) disease severity, and peripheral stress systems are frequently dysregulated in MS patients. However, the exact link between neurobiological stress systems and MS symptoms is unknown. To evaluate the link between neural stress responses and disease parameters, we used an arterial-spin-labeling functional MRI stress paradigm in 36 MS patients and 21 healthy controls. Specifically, we measured brain activity during a mental arithmetic paradigm with performance-adaptive task frequency and performance feedback and related this activity to disease parameters. Across all participants, stress increased heart rate, perceived stress, and neural activity in the visual, cerebellar and insular cortex areas compared with a resting condition. None of these responses was related to cognitive load (task frequency). Consistently, although performance and cognitive load were lower in patients than in controls, stress responses did not differ between groups. Insula activity elevated during stress compared with rest was negatively linked to impairment of pyramidal and cerebral functions in patients. Cerebellar activation was related negatively to gray matter (GM) atrophy (i.e., positively to GM volume) in patients. Interestingly, this link was also observed in overlapping areas in controls. Cognitive load did not contribute to these associations. The results show that our task induced psychological stress independent of cognitive load. Moreover, stress-induced brain activity reflects clinical disability in MS. Finally, the link between stress-induced activity and GM volume in patients and controls in overlapping areas suggests that this link cannot be caused by the disease alone.

Preserved brain metabolic activity at the age of 96 years.

Loss of brain tissue becomes notable to cerebral magnetic resonance imaging (MRI) at age 30 years, and progresses more rapidly from mid 60s. The incidence of dementia increases exponentially with age, and is all too frequent in the oldest old (≥ 90 years of age), the fastest growing age group in many countries. However, brain pathology and cognitive decline are not inevitable, even at extremely old age (den Dunnen et al., 2008).

Impulse control in the dorsolateral prefrontal cortex counteracts post-diet weight regain in obesity.

A variety of studies suggest that efficient treatments to induce short-term dietary success in obesity exist. However, sustained maintenance of reduced weight is rare as a large proportion of patients start to regain weight when treatment is discontinued. Thus, from a clinical perspective, it would be desirable to identify factors that counteract post-diet weight regain across longer time-scales. To address this question, we extended our previous work on neural impulse control mechanisms of short-term dietary success in obesity and now investigated the mechanisms counteracting long-term weight regain after a diet. Specifically, we measured neural impulse control during a delay discounting task with fMRI at two time points, i.e. the beginning ('T0') and the end ('T12') of a one-year follow-up interval after a 12-week diet. Then, we tested whether activity in the dorsolateral prefrontal cortex (DLPFC) at T0 and whether activity changes across the follow-up period (T0-T12) are linked to success in weight maintenance. The analyses conducted show that control-related DLPFC activity at T0 was coupled to the degree of success in weight maintenance. Consistently, also behavioral measures of control were linked to the degree of success in maintenance. A direct comparison of neural and behavioral control parameters for prognostic weight change modeling revealed that neural signals were more informative. Taken together, neural impulse control in the DLPFC measured with fMRI directly after a diet predicts real-world diet success in obese patients across extended time periods.

Performance reserves in brain-imaging-based phenotype prediction.

This study examines the impact of sample size on predicting cognitive and mental health phenotypes from brain imaging via machine learning. Our analysis shows a 3- to 9-fold improvement in prediction performance when sample size increases from 1,000 to 1 M participants. However, despite this increase, the data suggest that prediction accuracy remains worryingly low and far from fully exploiting the predictive potential of brain imaging data. Additionally, we find that integrating multiple imaging modalities boosts prediction accuracy, often equivalent to doubling the sample size. Interestingly, the most informative imaging modality often varied with increasing sample size, emphasizing the need to consider multiple modalities. Despite significant performance reserves for phenotype prediction, achieving substantial improvements may necessitate prohibitively large sample sizes, thus casting doubt on the practical or clinical utility of machine learning in some areas of neuroimaging.

Secure federated learning for Alzheimer's disease detection.

Machine Learning (ML) is considered a promising tool to aid and accelerate diagnosis in various medical areas, including neuroimaging. However, its success is set back by the lack of large-scale public datasets. Indeed, medical institutions possess a large amount of data; however, open-sourcing is prevented by the legal requirements to protect the patient's privacy. Federated Learning (FL) is a viable alternative that can overcome this issue. This work proposes training an ML model for Alzheimer's Disease (AD) detection based on structural MRI (sMRI) data in a federated setting. We implement two aggregation algorithms, Federated Averaging (FedAvg) and Secure Aggregation (SecAgg), and compare their performance with the centralized ML model training. We simulate heterogeneous environments and explore the impact of demographical (sex, age, and diagnosis) and imbalanced data distributions. The simulated heterogeneous environments allow us to observe these statistical differences' effect on the ML models trained using FL and highlight the importance of studying such differences when training ML models for AD detection. Moreover, as part of the evaluation, we demonstrate the increased privacy guarantees of FL with SecAgg via simulated membership inference attacks.

Benchmarking the influence of pre-training on explanation performance in MR image classification.

Convolutional Neural Networks (CNNs) are frequently and successfully used in medical prediction tasks. They are often used in combination with transfer learning, leading to improved performance when training data for the task are scarce. The resulting models are highly complex and typically do not provide any insight into their predictive mechanisms, motivating the field of "explainable" artificial intelligence (XAI). However, previous studies have rarely quantitatively evaluated the "explanation performance" of XAI methods against ground-truth data, and transfer learning and its influence on objective measures of explanation performance has not been investigated. Here, we propose a benchmark dataset that allows for quantifying explanation performance in a realistic magnetic resonance imaging (MRI) classification task. We employ this benchmark to understand the influence of transfer learning on the quality of explanations. Experimental results show that popular XAI methods applied to the same underlying model differ vastly in performance, even when considering only correctly classified examples. We further observe that explanation performance strongly depends on the task used for pre-training and the number of CNN layers pre-trained. These results hold after correcting for a substantial correlation between explanation and classification performance.

Evaluation of machine learning-based classification of clinical impairment and prediction of clinical worsening in multiple sclerosis.

BACKGROUND: Robust predictive models of clinical impairment and worsening in multiple sclerosis (MS) are needed to identify patients at risk and optimize treatment strategies. OBJECTIVE: To evaluate whether machine learning (ML) methods can classify clinical impairment and predict worsening in people with MS (pwMS) and, if so, which combination of clinical and magnetic resonance imaging (MRI) features and ML algorithm is optimal. METHODS: We used baseline clinical and structural MRI data from two MS cohorts (Berlin: n = 125, Amsterdam: n = 330) to evaluate the capability of five ML models in classifying clinical impairment at baseline and predicting future clinical worsening over a follow-up of 2 and 5 years. Clinical worsening was defined by increases in the Expanded Disability Status Scale (EDSS), Timed 25-Foot Walk Test (T25FW), 9-Hole Peg Test (9HPT), or Symbol Digit Modalities Test (SDMT). Different combinations of clinical and volumetric MRI measures were systematically assessed in predicting clinical outcomes. ML models were evaluated using Monte Carlo cross-validation, area under the curve (AUC), and permutation testing to assess significance. RESULTS: The ML models significantly determined clinical impairment at baseline for the Amsterdam cohort, but did not reach significance for predicting clinical worsening over a follow-up of 2 and 5 years. High disability (EDSS ≥ 4) was best determined by a support vector machine (SVM) classifier using clinical and global MRI volumes (AUC = 0.83 ± 0.07, p = 0.015). Impaired cognition (SDMT Z-score ≤ -1.5) was best determined by a SVM using regional MRI volumes (thalamus, ventricles, lesions, and hippocampus), reaching an AUC of 0.73 ± 0.04 (p = 0.008). CONCLUSION: ML models could aid in classifying pwMS with clinical impairment and identify relevant biomarkers, but prediction of clinical worsening is an unmet need.

Higher performance for women than men in MRI-based Alzheimer's disease detection.

INTRODUCTION: Although machine learning classifiers have been frequently used to detect Alzheimer's disease (AD) based on structural brain MRI data, potential bias with respect to sex and age has not yet been addressed. Here, we examine a state-of-the-art AD classifier for potential sex and age bias even in the case of balanced training data. METHODS: Based on an age- and sex-balanced cohort of 432 subjects (306 healthy controls, 126 subjects with AD) extracted from the ADNI data base, we trained a convolutional neural network to detect AD in MRI brain scans and performed ten different random training-validation-test splits to increase robustness of the results. Classifier decisions for single subjects were explained using layer-wise relevance propagation. RESULTS: The classifier performed significantly better for women (balanced accuracy [Formula: see text]) than for men ([Formula: see text]). No significant differences were found in clinical AD scores, ruling out a disparity in disease severity as a cause for the performance difference. Analysis of the explanations revealed a larger variance in regional brain areas for male subjects compared to female subjects. DISCUSSION: The identified sex differences cannot be attributed to an imbalanced training dataset and therefore point to the importance of examining and reporting classifier performance across population subgroups to increase transparency and algorithmic fairness. Collecting more data especially among underrepresented subgroups and balancing the dataset are important but do not always guarantee a fair outcome.

Predicting psychotic relapse following randomised discontinuation of paliperidone in individuals with schizophrenia or schizoaffective disorder: an individual participant data analysis.

BACKGROUND: Predicting relapse for individuals with psychotic disorders is not well established, especially after discontinuation of antipsychotic treatment. We aimed to identify general prognostic factors of relapse for all participants (irrespective of treatment continuation or discontinuation) and specific predictors of relapse for treatment discontinuation, using machine learning. METHODS: For this individual participant data analysis, we searched the Yale University Open Data Access Project's database for placebo-controlled, randomised antipsychotic discontinuation trials with participants with schizophrenia or schizoaffective disorder (aged ≥18 years). We included studies in which participants were treated with any antipsychotic study drug and randomly assigned to continue the same antipsychotic drug or to discontinue it and receive placebo. We assessed 36 prespecified baseline variables at randomisation to predict time to relapse, using univariate and multivariate proportional hazard regression models (including multivariate treatment group by variable interactions) with machine learning to categorise the variables as general prognostic factors of relapse, specific predictors of relapse, or both. FINDINGS: We identified 414 trials, of which five trials with 700 participants (304 [43%] women and 396 [57%] men) were eligible for the continuation group and 692 participants (292 [42%] women and 400 [58%] men) were eligible for the discontinuation group (median age 37 [IQR 28-47] years for continuation group and 38 [28-47] years for discontinuation group). Out of the 36 baseline variables, general prognostic factors of increased risk of relapse for all participants were drug-positive urine; paranoid, disorganised, and undifferentiated types of schizophrenia (lower risk for schizoaffective disorder); psychiatric and neurological adverse events; higher severity of akathisia (ie, difficulty or inability to sit still); antipsychotic discontinuation; lower social performance; younger age; lower glomerular filtration rate; benzodiazepine comedication (lower risk for anti-epileptic comedication). Out of the 36 baseline variables, predictors of increased risk specifically after antipsychotic discontinuation were increased prolactin concentration, higher number of hospitalisations, and smoking. Both prognostic factors and predictors with increased risk after discontinuation were oral antipsychotic treatment (lower risk for long-acting injectables), higher last dosage of the antipsychotic study drug, shorter duration of antipsychotic treatment, and higher score on the Clinical Global Impression (CGI) severity scale The predictive performance (concordance index) for participants who were not used to train the model was 0·707 (chance level is 0·5). INTERPRETATION: Routinely available general prognostic factors of psychotic relapse and predictors specific for treatment discontinuation could be used to support personalised treatment. Abrupt discontinuation of higher dosages of oral antipsychotics, especially for individuals with recurring hospitalisations, higher scores on the CGI severity scale, and increased prolactin concentrations, should be avoided to reduce the risk of relapse. FUNDING: German Research Foundation and Berlin Institute of Health.

Machine learning-based prediction of clinical outcomes after first-ever ischemic stroke.

Background: Accurate prediction of clinical outcomes in individual patients following acute stroke is vital for healthcare providers to optimize treatment strategies and plan further patient care. Here, we use advanced machine learning (ML) techniques to systematically compare the prediction of functional recovery, cognitive function, depression, and mortality of first-ever ischemic stroke patients and to identify the leading prognostic factors. Methods: We predicted clinical outcomes for 307 patients (151 females, 156 males; 68 ± 14 years) from the PROSpective Cohort with Incident Stroke Berlin study using 43 baseline features. Outcomes included modified Rankin Scale (mRS), Barthel Index (BI), Mini-Mental State Examination (MMSE), Modified Telephone Interview for Cognitive Status (TICS-M), Center for Epidemiologic Studies Depression Scale (CES-D) and survival. The ML models included a Support Vector Machine with a linear kernel and a radial basis function kernel as well as a Gradient Boosting Classifier based on repeated 5-fold nested cross-validation. The leading prognostic features were identified using Shapley additive explanations. Results: The ML models achieved significant prediction performance for mRS at patient discharge and after 1 year, BI and MMSE at patient discharge, TICS-M after 1 and 3 years and CES-D after 1 year. Additionally, we showed that National Institutes of Health Stroke Scale (NIHSS) was the top predictor for most functional recovery outcomes as well as education for cognitive function and depression. Conclusion: Our machine learning analysis successfully demonstrated the ability to predict clinical outcomes after first-ever ischemic stroke and identified the leading prognostic factors that contribute to this prediction.

Similar neural pathways link psychological stress and brain-age in health and multiple sclerosis.

Clinical and neuroscientific studies suggest a link between psychological stress and reduced brain health in health and neurological disease but it is unclear whether mediating pathways are similar. Consequently, we applied an arterial-spin-labeling MRI stress task in 42 healthy persons and 56 with multiple sclerosis, and investigated regional neural stress responses, associations between functional connectivity of stress-responsive regions and the brain-age prediction error, a highly sensitive machine learning brain health biomarker, and regional brain-age constituents in both groups. Stress responsivity did not differ between groups. Although elevated brain-age prediction errors indicated worse brain health in patients, anterior insula-occipital cortex (healthy persons: occipital pole; patients: fusiform gyrus) functional connectivity correlated with brain-age prediction errors in both groups. Finally, also gray matter contributed similarly to regional brain-age across groups. These findings might suggest a common stress-brain health pathway whose impact is amplified in multiple sclerosis by disease-specific vulnerability factors.

Uncontrolled eating and sensation-seeking partially explain the prediction of future binge drinking from adolescent brain structure.

Binge drinking behavior in early adulthood can be predicted from brain structure during early adolescence with an accuracy of above 70%. We investigated whether this accurate prospective prediction of alcohol misuse behavior can be explained by psychometric variables such as personality traits or mental health comorbidities in a data-driven approach. We analyzed a subset of adolescents who did not have any prior binge drinking experience at age 14 (IMAGEN dataset, n = 555, 52.61% female). Participants underwent structural magnetic resonance imaging at age 14, binge drinking assessments at ages 14 and 22, and psychometric questionnaire assessments at ages 14 and 22. We derived structural brain features from T1-weighted magnetic resonance and diffusion tensor imaging. Using Machine Learning (ML), we predicted binge drinking (age 22) from brain structure (age 14) and used counterbalancing with oversampling to systematically control for 110 + variables from a wide range of social, personality, and other psychometric characteristics potentially associated with binge drinking. We evaluated if controlling for any variable resulted in a significant reduction in ML prediction accuracy. Sensation-seeking (-13.98 ± 1.68%), assessed via the Substance Use Risk Profile Scale at age 14, and uncontrolled eating (-13.98 ± 3.28%), assessed via the Three-Factor-Eating-Questionnaire at age 22, led to significant reductions in mean balanced prediction accuracy upon controlling for them. Thus, sensation-seeking and binge eating could partially explain the prediction of future binge drinking from adolescent brain structure. Our findings suggest that binge drinking and binge eating at age 22 share common neurobiological precursors discovered by the ML model. These neurobiological precursors seem to be associated with sensation-seeking at age 14. Our results facilitate early detection of increased risk for binge drinking and inform future clinical research in trans-diagnostic prevention approaches for adolescent alcohol misuse.

Prediction of high and low disease activity in early MS patients using multiple kernel learning identifies importance of lateral ventricle intensity.

Background: Lack of easy-to-interpret disease activity prediction methods in early MS can lead to worse patient prognosis. Objectives: Using machine learning (multiple kernel learning - MKL) models, we assessed the prognostic value of various clinical and MRI measures for disease activity. Methods: Early MS patients (n = 148) with at least two associated clinical and MRI visits were investigated. T2-weighted MRIs were cropped to contain mainly the lateral ventricles (LV). High disease activity was defined as surpassing NEDA-3 Criteria more than once per year. Clinical demographic, MRI-extracted image-derived phenotypes (IDP), and MRI data were used as inputs for separate kernels to predict future disease activity with MKL. Model performance was compared using bootstrapped effect size analysis of mean differences. Results: A total of 681 visits were included, where 81 (55%) patients had high disease activity in a combined end point measure using all follow-up visits. MKL model discrimination performance was moderate (AUC ≥ 0.62); however, modelling with combined clinical and cropped LV kernels gave the highest prediction performance (AUC = 0.70). Conclusions: MRIs contain valuable information on future disease activity, especially in and around the LV. MKL techniques for combining different data types can be used for the prediction of disease activity in a relatively small MS cohort.

Nucleus basalis of Meynert predicts cognition after deep brain stimulation in Parkinson's disease.

INTRODUCTION: Subthalamic DBS in Parkinson's disease has been associated with cognitive decline in few cases. Volume reduction of the nucleus basalis of Meynert (NBM) seems to precede cognitive impairment in Parkinson's disease. In this retrospective study, we evaluated NBM volume as a predictor of cognitive outcome 1 year after subthalamic DBS. METHODS: NBM volumes were calculated from preoperative MRIs using voxel-based morphometry. Cognitive outcome was defined as the relative change of MMSE or DemTect scores from pre-to 1 year postoperatively. A multiple linear regression analysis adjusted for the number of cognitive domains affected in the preoperative neuropsychological testing and UPDRS III was conducted. To account for other variables and potential non-linear effects, an additional machine learning analysis using random forests was applied. RESULTS: 55 patients with Parkinson's disease (39 male, age 61.4 ± 7.5 years, disease duration 10.8 ± 4.7 years) who received bilateral subthalamic DBS electrodes at our center were included. Although overall cognition did not change significantly, individual change in cognitive abilities was variable. Cognitive outcome could be predicted based on NBM size (B = 208.98, p = 0.022*) in the regression model (F(3,49) = 2.869; R2 of 0.149; p = 0.046*). Using random forests with more variables, cognitive outcome could also be predicted (average root mean squared error between predicted and true cognitive change 11.28 ± 9.51, p = 0.039*). Also in this model, NBM volume was the most predictive variable. CONCLUSION: NBM volume can be used as a simple non-invasive predictor for cognitive outcome after DBS in Parkinson's disease, especially when combined with other clinical parameters that are prognostically relevant.

Generating 3D TOF-MRA volumes and segmentation labels using generative adversarial networks.

Deep learning requires large labeled datasets that are difficult to gather in medical imaging due to data privacy issues and time-consuming manual labeling. Generative Adversarial Networks (GANs) can alleviate these challenges enabling synthesis of shareable data. While 2D GANs have been used to generate 2D images with their corresponding labels, they cannot capture the volumetric information of 3D medical imaging. 3D GANs are more suitable for this and have been used to generate 3D volumes but not their corresponding labels. One reason might be that synthesizing 3D volumes is challenging owing to computational limitations. In this work, we present 3D GANs for the generation of 3D medical image volumes with corresponding labels applying mixed precision to alleviate computational constraints. We generated 3D Time-of-Flight Magnetic Resonance Angiography (TOF-MRA) patches with their corresponding brain blood vessel segmentation labels. We used four variants of 3D Wasserstein GAN (WGAN) with: 1) gradient penalty (GP), 2) GP with spectral normalization (SN), 3) SN with mixed precision (SN-MP), and 4) SN-MP with double filters per layer (c-SN-MP). The generated patches were quantitatively evaluated using the Fréchet Inception Distance (FID) and Precision and Recall of Distributions (PRD). Further, 3D U-Nets were trained with patch-label pairs from different WGAN models and their performance was compared to the performance of a benchmark U-Net trained on real data. The segmentation performance of all U-Net models was assessed using Dice Similarity Coefficient (DSC) and balanced Average Hausdorff Distance (bAVD) for a) all vessels, and b) intracranial vessels only. Our results show that patches generated with WGAN models using mixed precision (SN-MP and c-SN-MP) yielded the lowest FID scores and the best PRD curves. Among the 3D U-Nets trained with synthetic patch-label pairs, c-SN-MP pairs achieved the highest DSC (0.841) and lowest bAVD (0.508) compared to the benchmark U-Net trained on real data (DSC 0.901; bAVD 0.294) for intracranial vessels. In conclusion, our solution generates realistic 3D TOF-MRA patches and labels for brain vessel segmentation. We demonstrate the benefit of using mixed precision for computational efficiency resulting in the best-performing GAN-architecture. Our work paves the way towards sharing of labeled 3D medical data which would increase generalizability of deep learning models for clinical use.

Structural differences in adolescent brains can predict alcohol misuse.

Alcohol misuse during adolescence (AAM) has been associated with disruptive development of adolescent brains. In this longitudinal machine learning (ML) study, we could predict AAM significantly from brain structure (T1-weighted imaging and DTI) with accuracies of 73 -78% in the IMAGEN dataset (n∼1182). Our results not only show that structural differences in brain can predict AAM, but also suggests that such differences might precede AAM behavior in the data. We predicted 10 phenotypes of AAM at age 22 using brain MRI features at ages 14, 19, and 22. Binge drinking was found to be the most predictable phenotype. The most informative brain features were located in the ventricular CSF, and in white matter tracts of the corpus callosum, internal capsule, and brain stem. In the cortex, they were spread across the occipital, frontal, and temporal lobes and in the cingulate cortex. We also experimented with four different ML models and several confound control techniques. Support Vector Machine (SVM) with rbf kernel and Gradient Boosting consistently performed better than the linear models, linear SVM and Logistic Regression. Our study also demonstrates how the choice of the predicted phenotype, ML model, and confound correction technique are all crucial decisions in an explorative ML study analyzing psychiatric disorders with small effect sizes such as AAM.

Patch individual filter layers in CNNs to harness the spatial homogeneity of neuroimaging data.

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.