Schedule

In this two-day workshop we will explore perspectives, novel methodology, and its impact at the interface of Machine Learning, Inference, Neuroimaging and Neuroscience. We aim to bring researchers from the machine learning and neuroscience communities together, in order to discuss open questions, identify the core points for a number of the controversial issues, and discuss the interaction among novel methodology and neuroscientific problems.

The workshop will be structured around 3 main topics:
  • machine learning and pattern recognition methodology
  • causal inference in neuroimaging
  • linking machine learning, neuroimaging and neuroscience
Each session will be opened by 1-2 invited talks, and an in depth discussion. This will be followed by original contributions. Original contributions will also be presented and discussed during a poster session. Each workshop day will end with a panel discussion, during which we will address specific questions, and invited speakers will open each segment with a brief presentation of their opinion.

Location

  • Workshop Number: 141
  • Room: Harveys Emerald Bay 5

Friday, December 7, 2012

Morning session: 7:30-10:30

Contributed talks: 20 minutes talk + 5 minutes questions

Invited talks: 40 minute talk + 10 minute questions

 7:30-7:40Opening remarks
 7:40-8:30Invited talk - Klaus-Robert Müller: Decoding Cognitive States with BCI
 8:30-8:55Jessica ThompsonMusical neurosemantic decoding using online weighted approximate-rank pairwise loss optimization in a joint semantic space
 8:55-9:15Coffee break/poster setup (day 1 posters)
 9:15-9:40Leila Wehbe: Tracking Story Reading in the Brain
 9:40-10:00Discussion
 10:00-10:30

Poster session (day 1 posters), continued over lunch break

1. Discriminant BOLD Activation Patterns during Mental Imagery in Parkinson's Disease

2. Mining the brain with a theory of visual attention

3. Decoding Word Semantics from Magnetoencephalography Time Series Transformations

4. Multi-scale automated cell segmentation in two-photon calcium imaging

5. Learning Latent Structure for Identifying Neuronal Activity

6.  Fast Simultaneous Training of Generalized Linear Models (FaSTGLZ) for Multi-voxel Pattern Analysis in fMRI

7. Musical neurosemantic decoding using online weighted approximate-rank pairwise loss optimization in a joint semantic space

8. Tracking Story Reading in the Brain

9. Probabilistic M/EEG source imaging from sparse spatio-temporal event structure

Afternoon session: 3:30-6:30

 3:30-4:20Invited talk - Martin Lindquist: Connectivity and Causality in Brain Imaging
 4:20-4:45Carsten Stahlhut: Probabilistic M/EEG source imaging from sparse spatio-temporal event structure
 4:45-5:10Coffee break/poster session (day 1 posters)
 5:10-6:00Invited talk - Jack Gallant: Voxel-wise modeling and decoding (VWMD): Opportunities and challenges
 6:00-6:30Panel discussion

Saturday, December 8, 2012

Morning session: 7:30-10:30

 7:30-8:20Invited talk - Francisco Pereira: Towards a generic pattern-based analysis of multi-class fMRI data
 8:20-8:45Christopher Baldessano: Discovering Voxel-Level Functional Connectivity Between Cortical Regions
 8:45-9:15Coffee break/poster setup (day 2 posters)
 9:15-9:40Karl-Heinz Nenning: Overlapping Functional Networks in Multiple fMRI Paradigms
 9:40-10:00Discussion
 10:00-10:30

Poster session (day 2 posters), continued over lunch break

10. Factor Topographic Latent Source Analysis: Factor Analysis for Brain Images

11. Relating Structural MRI, Demographic and Cognitive Ability Measures to Functional MRI Measures

12. Elastic-net Multiple Kernel Learning for multi-region neuroimaging based diagnosis

13. Joint Modelling of Structural and Functional Brain Networks

14.  EEG-based single-trial classification of targets in a rapid serial visual presentation task with supervised spatial filtering

15. OASIS is Automated Statistical Inference for Segmentation with applications to multiple sclerosis lesion segmentation in MRI

16. Discovering Voxel-Level Functional Connectivity Between Cortical Regions

17. Overlapping Functional Networks in Multiple fMRI Paradigms

18. On the Interpretability of Linear Multivariate Neuroimaging Analyses: Filters, Patterns and their Relationship


Afternoon session: 3:30-6:30

 3:30-4:20Invited talk - Mert Sabuncu: The Relevance Voxel Machine: Bayesian image-based prediction
 4:20-4:45Felix Biessmann: On the Interpretability of Linear Multivariate Neuroimaging Analyses: Filters, Patterns and their Relationship
 4:45-5:10Coffee break/poster session (day 2 posters)
 5:10-6:00Invited talk - Bertrand Thirion: Spatial regularization and sparsity for brain mapping
 6:00-6:30Panel discussion

Invited Talks:

  • Jack Gallant (UC Berkeley): Voxel-wise modeling and decoding (VWMD): Opportunities and challenges

    Abstract: Many modeling techniques have been developed to try to recover the maximum available information from fMRI data. One approach that has become increasingly popular is voxel-wise modeling and decoding (VWMD). Under this approach each voxel is modeled separately by estimating the transformation between a set of hypothetical stimulus or task features and measured BOLD responses. Different hypothetical feature spaces can be evaluated by comparing model predictions; feature tuning in individual voxels can be assessed; and tuning patterns across voxels can be determined by clustering or dimensionality reduction. Although VWMD has certain similarities to other pattern analysis techniques used in fMRI, it does have a few specific advantages over competing approaches. Most importantly, it permits analysis of complex, continuous experimental designs that would be extremely difficult to assess using other methods. In this talk I will review VWMD, address factors that have a significant effect on model quality, and discuss some of the remaining challenges for the approach.

  • Martin Lindquist (John Hopkins U.): Connectivity and Causality in Brain Imaging
Abstract: To date human brain mapping has primarily been used to construct maps indicating regions of the brain that are activated by certain tasks. Recently, there has been an increased interest in augmenting this type of analysis with connectivity studies that seek to describe how brain regions interact and how these interactions depend on experimental conditions and behavioral measures. Often researchers discriminate between functional connectivity, the undirected association between two or more fMRI time series, and effective connectivity, the directed influence of one brain region on the physiological activity recorded in other brain regions. In this talk we argue that this distinction is not entirely clear or relevant. Instead, the validity of the conclusions made from any connectivity method will depend strongly on certain key assumptions which are often poorly specified and difficult to check. We illustrate how ideas from causal inference can provide a mathematical framework for determining these assumptions. We conclude with the introduction of a functional path analysis model for studying brain connectivity, which extends the standard structural equation model framework to the functional data setting.
  • Francisco Pereira (Princeton): Towards a generic pattern-based analysis of multi-class fMRI data

In this talk I will describe some of the tools we have been developing for searchlight-style similarity and classification analyses, and our work in progress to compose them into a generic method for extracting the different kinds of information present in a multi-class fMRI dataset.

  • Mert Sabuncu (MGH):  The Relevance Voxel Machine: Bayesian image-based prediction            

Abstract: The Relevance Voxel Machine (RVoxM) is a dedicated Bayesian model for making predictions based on medical imaging data. In contrast to the generic machine learning algorithms that have often been used for this purpose, the method is designed to utilize a small number of spatially clustered sets of voxels that are particularly suited for clinical interpretation. RVoxM automatically tunes all its free parameters during the training phase, and offers the additional advantage of producing probabilistic prediction outcomes. We demonstrate RVoxM as a regression model by predicting age from volumetric gray matter segmentations, and as a classification model by distinguishing patients with Alzheimer’s disease from healthy controls using surface-based cortical thickness data. Our results indicate that RVoxM yields biologically meaningful models, while providing state-of-the-art predictive accuracy.

  • Bertrand Thirion: Spatial regularization and sparsity for brain mapping

    Abstract: Brain activity decoding, is a recent paradigm for analyzing functional magnetic resonance imaging (fMRI) data, based on pattern recognition tools. It enables neuroscientists to take into account the multivariate information between voxels and is currently the only way to assess how precisely some cognitive information is encoded by the activity of neural populations within the whole brain. However, it relies on a prediction function that is plagued by the curse of dimensionality, as we have far more features than samples, i.e., more voxels than fMRI volumes.
    While many machine learning solutions have been proposed to deal with this kind of issues, with more or less sophistication, these techniques are designed to maximize prediction accuracy, thus solving a supervised learning problem. It turns out that neuroimagers are often interested in interpreting the spatial pattern used by the model, a question to which popular machine learning algorithms do not give a satisfactory response: on the theoretical side, many methods do not offer any guarantee to pick a good model, or even to converge to the right model with many observations; in practice, many different models relying on the same spatial pattern can produce strikingly different solutions. We will first review some key concepts of pattern analysis, such as regularization and sparsity, and some known results on the quality of spatial pattern recovery in high-dimensional prediction problems. Then we will discuss recent solutions proposed to deal with this problem, essentially trying to build discriminative patterns based on natural hypotheses of the problem: clustering and spatial smoothness.
  • Klaus-Robert Müller (TU-Berlin): Decoding Cognitive States with BCI

The last years have seen a rise in interest in using BCI methodology for investigating non-medical questions beyond the purpose of communication and control. The first part of the talk gives a short introduction to BCI challenges from a machine learning perspective.  In the remaining 2 sections of the talk we present our work on selected applications of BCI. The main part of this talk will discuss the use of EEG in combination with BCI methods for investigating how signal quality is processed (non-)consciously on a neural level (auditory/visual domain).

If time permits, in the third part of this talk we present results of an EEG study on forced emergency braking in a driving simulator. Participants had to closely follow a computer-controlled lead vehicle, which would induce emergency situations by braking abruptely in random intervals. Grand-average results (N=18) indicate that interpretable neural signatures as well as (to some extent) muscle activity at the lower leg allow for much earlier detection of the driver's intention to perform emergency braking than behavioural measures, while achieving the same detection accuracy.

This is joint work with members of the bbci.de team and in particular
my collegues Blankertz, Curio, Haufe, Wiegand, Scholler, and Porbadnigk.


Accepted Papers:

Oral presentations:

Title: On the Interpretability of Linear Multivariate Neuroimaging Analyses: Filters, Patterns and their Relationship
Abstract: Multivariate linear methods are an important tool for decoding neural sources from neuroimaging data. However results obtained from multivariate linear methods are not as easy to interpret as results from mass-univariate analyses. A common misconception is that multivariate filters, that transform measured signals into neural sources of interest, can be interpreted as activation patterns, which reflect underlying neural processes. Yet filters and patterns are not the same. This paper tries to create some awareness for the difference between multivariate filters and activation patterns. We show the difference in simulations and real data examples and recapitulate a simple but efficient way of transforming one into the other.
Author Names: Felix Biessmann*, TU Berlin, Abt. ML, FR 6-9
Sven Daehne, TU Berlin, Abt. ML, FR 6-9
Frank Meinecke, TU Berlin, Dept. ML
Kai Goergen, TU Berlin, Dept. ML
Benjamin Blankertz, TU Berlin, Dept. ML
Stefan Haufe, TU Berlin, Dept. ML


Title: Discovering Voxel-Level Functional Connectivity Between Cortical Regions
Abstract: Functional connectivity patterns are known to exist in the human brain at the millimeter scale, but the standard fMRI connectivity measure only computes functional correlations at a coarse level. We present a method for identifying fine-grained functional connectivity between any two brain regions by simultaneously learning voxel-level connectivity maps over both regions. We show how to formulate this problem as a constrained least-squares optimization, which can be solved using a trust region approach. Our method can automatically discover multiple correspondences between distinct voxel clusters in the two regions, even when these clusters have correlated timecourses. We validate our method by identifying a known division in the lateral occipital complex using only functional connectivity, thus demonstrating that we can successfully learn subregion connectivity structures from a small amount of training data.
Author Names: Christopher Baldassano*, Stanford University
Marius Cătălin Iordan, Stanford University
Diane Beck, University of Illinois at Urbana-Champaign
Li Fei-Fei, Stanford University

Title: Overlapping Functional Networks in Multiple fMRI Paradigms
Abstract: Studying the modular composition of the functional cerebral architecture is a challenging line of research. Of particular interest are network structures that are active during specific cognitive tasks. In this paper we address the question of identifying partially overlapping networks that are active across different fMRI experiment conditions. We propose to use Multiple Relational Embedding (MRE) on functional brain imaging data acquired during different cognitive tasks. Multiple functional relationships are embedded into a single joint latent embedding, that encodes both joint, and individual network structure.
Author Names: Karl-Heinz Nenning*, CIR Lab, Medical University of Vienna
Georg Langs, CIR Lab, Medical University of Vienna

Title: Musical neurosemantic decoding using  online weighted approximate-rank pairwise loss optimization in a joint semantic space

Abstract: The field of neurosemantics is concerned with the mapping between concepts and the neural activity that they elicit. Automatic media annotation, similarly, seeks to label media with meaningful or perceptually relevant words. These annotations (a.k.a. tags) are typically high-level, human judgments that allow us to organize and retrieve documents according to conceptual search queries. In the present work, an algorithm originally designed for automatic media annotation was adapted for musical neurosemantic decoding. This adapted algorithm, which we call MUNSE (Music Understanding by NeuroSemantic Embedding), was used to predict musical tags from hemodynamic brain activity evoked by music. MUNSE was compared to several baseline measures by calculating precision@n for n ∈ {1, 3, 5, 10}. MUNSE performed significantly above chance across all subjects and for all n. MUNSE outperformed KNN for large n but not for small n. This suggests that the neural encoding of the highest ranked tags is dominant and consistent enough that it can be captured using simple similarity measures. Our method, however, is able to accurately predict more lower-ranked tags and thus capture more nuanced information about the neural encoding of the music. We suggest that such a hybrid approach of neurologically informed information re-trieval could be used to improve performance on tasks related to human cognition, e.g. automatic semantic labeling.
Author Names: Jessica Thompson*, Dartmouth College
Michael Casey,
Lorenzo Torresani

Title: Tracking Story Reading in the Brain
Abstract: Our research goal is to study information processing in the brain during story comprehension. Story comprehension is a complex task that involves many simultaneous levels of processing. We present a method that models the story as a time varying series of visual, semantic, syntactic and discourse features, and maps these features to their neural representation. We test the model by computing the accuracy with which it can predict the passage that is being read from a segment of fMRI data. Our result show that it performs significantly better than chance. Furthermore, we find a different pattern of representation in the brain for every type of feature.
Author Names: Leila Wehbe*, Carnegie Mellon University
Partha Talukdar, Carnegie Mellon University
Brian Murphy, CMU
Alona Fyshe, Carnegie Mellon University
Gustavo Sudre, Carnegie Mellon University
Tom Mitchell, Carnegie Mellon University

Title: Probabilistic M/EEG source imaging from sparse spatio-temporal event structure
Abstract: While MEG and EEG source imaging methods have to tackle a severely ill-posed problem their success can be stated as their ability to constrain the solutions using appropriate priors. In this paper we propose a hierarchical Bayesian model facilitating spatio-temporal patterns through the use of both spatial and temporal basis functions. We demonstrate the efficacy of the model on both artificial data and real EEG data.
Author Names: Carsten Stahlhut*, DTU Informatics
Hagai Attias, Convex Imaging
David Wipf, Microsoft Research Asia
Lars Hansen, Technical University of Denmark
Srikantan Nagarajan, University of California, San Francisco



Posters:


Friday:


 
Title: Discriminant BOLD Activation Patterns during Mental Imagery in Parkinson's Disease
Abstract: Using machine learning based models in clinical applications has become current practice and can prove useful to provide information at the subject's level, such as predicting an (early) diagnosis or monitoring the evolution of a disease. However, the performance of these models depends on the choice of a biomarker to detect the presence or absence of a disease. Choosing a biomarker is not straightforward, especially in the case of Parkinson's disease when compared to healthy subjects. In the present work, we investigated the mental imagery of gait as a biomarker of Parkinson's disease and showed that the signal in the mesencephalic locomotor region during the mental imagery of gait at a comfortable pace can discriminate significantly between idiopathic Parkinson's disease patients and healthy subjects. Although there is room for improvement, the results of this preliminary study are promising.
Author Names: Jessica Schrouff*, Cyclotron Research Centre
Julien Cremers, Cyclotron Research Centre
Kevin D' Ostilio, Cyclotron Research Centre
Valérie Delvaux, Department of Neurology, Liège University Hospital, Belgium
Gaëtan Garraux, Cyclotron Research Centre
Christophe Phillips, Cyclotron Research Centre

Title: Mining the brain with a theory of visual attention
Abstract: We present a new supervised component analysis method with an application to EEG. The method detects and extracts compo- nents that are predictive of behavior relative to an expected value which is derived from a formal psychological theory of visual attention. We an- alyze the pre-stimulus EEG activity from a single-letter recognition task and find distinct components that each contribute to the joint prediction through separable perceptual parameters on the single-trial level.
Author Names: Mads Dyrholm*, University of Copenhagen
Maria Nordfang,
Claus Bundesen

Title: Decoding Word Semantics from Magnetoencephalography Time Series Transformations
Abstract: Neuroimaging techniques such as Magnetoencephalography have facilitated the careful study of perceptual and motor systems. These processes are largely feed-forward and bottom up, and the evoked responses are very consistent. Gaining a similarly strong understanding of higher-level cognitive thought has proven more difficult. The processes involved in higher-order thought appear to be spatially distributed, involve top-down cognitive influence and are not as tightly coupled to the stimulus. To deal with these complications and inconsistencies, we need a robust method for processing the MEG signal. In this study we explore several methods of processing the MEG signal and evaluate their utility for decoding the higher order cognitive process of noun comprehension.
Author Names: Alona Fyshe*, Carnegie Mellon University
Gustavo Sudre, Carnegie Mellon University
Leila Wehbe, Carnegie Mellon University
Brian Murphy, CMU
Tom Mitchell, Carnegie Mellon University

Title: Multi-scale automated cell segmentation in two-photon calcium imaging
Abstract: Two photon calcium imaging (TPCI) is a relatively new and very promising technique for in vivo imaging of the structure and func- tion of neural populations. However, the data processing methodology for TPCI is underdeveloped. This presents an opportunity for statistics and machine learning to contribute substantively to basic neuroscience by providing a principled analysis pipeline that can be used by experimenters. We present here a procedure for automating the detection of cells in TPC images. Our procedure consists of an unsupervised multi-scale blob detector to generate candidate cell masks, and a minimally supervised Random Forest classifier to verify or discard the candidates. Our procedure improves on existing techniques, requiring minimal supervision, adapting to spatial inhomogeneities, and applying generally over cell types and animals.
Author Names: Bronwyn Woods*, Carnegie Mellon University
Alberto Vazquez,
Seong-Gi Kim,
William Eddy,

Title: Learning Latent Structure for Identifying Neuronal Activity
Abstract:
In the mammalian hippocampus, spatio-temporal co--activation of neurons are crucial elements for the formation of spatial memories; thus being important reliable monitoring and detection of distributed patterns of activity at single--neuron resolution. Traditionally, the neuronal activity patterns are detected through micro--electrode recordings but are not able to reveal the spatial pattern activity. Hence, optical imaging techniques based on activity--dependent intracellular Ca2+-signals have been recently emerged to observe these spatial patterns; however existing methods based on regions of interest (ROI) defined manually or simple component analysis with a subsequent image segmentation only aim to detect neuronal activity of single cells since the co--activation patterns are identified based on heuristics. Therefore, in this paper, we formulate the identification of neuronal activity of single cells and the neuronal co--activation into the same framework. Specifically, we aim to learn a latent structure that relates between co--activation patterns and single neurons; thus identifying and monitoring the activation of neurons at different levels, i.e. single neurons or group of co--activated neurons, at the same time. This provides a tool to investigate cognition--related activity patterns and fill the gap between different levels of neuronal activity representation.
Author Names: Ferran Diego*, HCI, University of Heidelberg
Fred Hamprecht,

Title: Fast Simultaneous Training of Generalized Linear Models (FaSTGLZ) for Multi-voxel Pattern Analysis in fMRI
Abstract: We present an efficient algorithm for simultaneously training elastic-net-regularized generalized linear models across many related problems, which may arise from bootstrapping, cross-validation and nonparametric permutation testing. Our approach leverages the redundancies across problems to obtain $\approx 10$x computational improvements relative to solving the problems sequentially by the glmnet algorithm of \cite{Friedman10}. We demonstrate our fast simultaneous training of generalized linear models (FaSTGLZ) algorithm, for multivariate analysis of fMRI and run otherwise computationally intensive bootstrapping and permutation test analyses that are typically necessary for obtaining statistically rigorous classification results and meaningful interpretation.
Author Names: Bryan Conroy*, Columbia University
Jennifer Walz, Columbia University
Paul Sajda, Columbia University

Saturday:

Title:


Factor Topographic Latent Source Analysis: Factor Analysis for Brain Images
Abstract: Traditional approaches to analyzing experimental functional magnetic resonance imaging (fMRI) data entail fitting per-voxel parameters to explain how the observed images reflect the thoughts and stimuli a participant experienced during the experiment. These methods implicitly assume that voxel responses are independent and that the unit of analysis should be the voxel. However, both of these assumptions are known to be untrue: it is well known that voxel activations exhibit strong spatial correlations, and common sense tells us that the true underlying brain activations are independent of the resolution at which the brain image happened to be taken. Here we propose a fundamentally different approach, whereby brain images are represented as weighted sums of spatial functions. Our technique yields compact representations of the brain images that leverage spatial correlations in the data and are independent of the image resolution.
Author Names: Jeremy Manning*, Princeton University
Kenneth Norman, Princeton University
David Blei, Princeton University


Title: Relating Structural MRI, Demographic and Cognitive Ability Measures to Functional MRI Measures
Abstract: A select combination of structural magnetic resonance imaging (MRI), demographic, and/or cognitive ability measures in the elderly have been shown to be associated with functional MRI measures. The goal of this study is to identify the relative strength of the different structural imaging, demographic and cognitive ability features at predicting functional connectivity at rest. This study uses both linear regression and artificial neural networks to achieve this goal. The results suggest that linear regression performs better and the features that best relate to functional measures include age, mini-mental state examination scores, total regional volume, number of tracks connecting regions, and regional white matter hyperintensities volume.
Author Names: Meenal Patel*, University of Pittsburgh
Howard Aizenstein, University of Pittsburgh

Title: Elastic-net Multiple Kernel Learning for multi-region neuroimaging based diagnosis
Abstract: Multiple Kernel Learning (MKL) has been proposed as an approach to simultaneously learn the kernel weights and the associated decision function in supervised learning settings (e.g. [1]). In the context of neuroimaging-based classification MKL framework can be applied to investigate the relative contributions of different image modalities and/or brain regions. Here we applied the Elastic-net MKL as an exploratory approach to learn the optimal combination of individual brain regions that best discriminate depressed patients versus healthy subjects based on patterns of fMRI activation to sad faces. Our results show that the application of Elastic-net MKL to neuroimaging data can lead to anatomically interpretable classification models and potentially improve the accuracy with respect to the whole-brain accuracy. In addition it can provide insights about how a psychiatric or neurologic disorder affects the brain, i.e. sparse vs. distributed effects.
Author Names: Janaina Mourao-Miranda*, University College London
Jane Rondina, University College London
Liana Portugal, University College London
John Shawe-Taylor, University College London

Title: Joint Modelling of Structural and Functional Brain Networks
Abstract: Functional and structural magnetic resonance imaging have become the most important noninvasive windows to the human brain. A major challenge in the analysis of brain networks is to establish the similarities and dissimilarities between functional and structural connectivity. We formulate a non-parametric Bayesian network model which allows for joint modelling and integration of multiple networks. We demonstrate the model's ability to detect vertices that share structure across networks jointly in functional MRI (fMRI) and diffusion MRI (dMRI) data. Using two fMRI and dMRI scans per subject, we establish significant structures that are consistently shared across subjects and data splits. This provides an unsupervised approach for modeling of structure-function relations in the brain and provides a general framework for multimodal integration.
Author Names: Kasper Andersen*, Technical University of Denmar
Tue Herlau, Technical University of Denmark
Morten Mørup, Technical University of Denmark
Mikkel Schmidt, Technical University of Denmark
Kristoffer Madsen, Danish Research Centre for Magnetic Resonance
Mark Lyksborg, Technical University of Denmark
Tim Dyrby, Danish Research Centre for Magnetic Resonance
Hartwig Siebner, Danish Research Centre for Magnetic Resonance
Lars Hansen, Technical University of Denmark

Title: EEG-based single-trial classification of targets in a rapid serial visual presentation task with supervised spatial filtering
Abstract: The detection of single-trial event related potentials (ERPs) in the electroencephalogram (EEG) is a difficult problem because of the poor spatial resolution and low signal-to-noise ratio of the EEG signal. Efficient signal processing and machine learning techniques have overcome this problem thanks to spatial filtering techniques, which have been used to enhance the relevant information in the signal by combining the signal recorded across the different sensors. We propose a neural network with a convolutional layer dedicated to spatial filtering (CNN) for the detection of ERPs during a rapid serial visual presentation (RSVP) task. The method is compared with two state of the art methods for supervised spatial filtering (xDAWN and Common Spatial Patterns). The two latter methods are combined with a neural network (MLP) for the classification. We have compared these methods with an MLP without spatial filtering as pre-processing. These techniques were evaluated on an RSVP task where eight participants had to detect faces from car images. Whereas the best performance is obtained with CNN, with an area under the ROC curve (AUC) of 0.843, the MLP reaches an AUC of 0.820.
Author Names: Hubert Cecotti*, University of California
Miguel Eckstein,
Barry Giesbrecht

Title: OASIS is Automated Statistical Inference for Segmentation with applications to multiple sclerosis lesion segmentation in MRI
Abstract: Magnetic resonance imaging (MRI) can be used to detect lesions in the brains of multiple sclerosis (MS) patients and is essential for evaluating disease-modifying therapies and monitoring disease progression. We propose OASIS is Automated Statistical Inference for Segmentation (OASIS), an automated statistical method for segmenting MS lesions in MRI studies. We use logistic regression models incorporating multiple MRI modalities to estimate voxel-level probabilities of lesion presence. Intensity-normalized T1-weighted, T2-weighted, fluid-attenuated inversion recovery and proton density volumes from 131 MRI studies (98 MS subjects, 33 healthy subjects) with manual lesion segmentations are used to train and validate our model. Within this set, OASIS detected lesions with an area under the receiver-operator characteristic curve of 98% (95% CI; [96%, 99%]) at the voxel level. Use of intensity-normalized MRI volumes enables OASIS to be robust to variations in scanners and acquisition sequences. To further validate the method, we applied OASIS to 169 MRI studies acquired at a separate imaging center. An experienced MS neuroradiologist compared these segmentations to those produced by LesionTOADS, an image segmentation software that provides segmentation of both lesions and normal brain structures. For lesions, OASIS out-performed LesionTOADS in 77% (95% CI: [71%, 83%]) of cases. For a randomly selected subset of 50 of these studies, one additional radiologist and one neurologist also scored the images. Within this set, the neuroradiologist ranked OASIS higher than LesionTOADS in 76% (95% CI: [64%, 88%]) of cases, the neurologist 66% (95% CI: [52%, 78%]) and the radiologist 52% (95% CI: [38%, 66%]).
Author Names: Elizabeth Sweeney*, Johns Hopkins
Russell Shinohara, University of Pennsylvania
Navid Shiee,
Joshua Vogelstein,
Farrah Mateen,
Avni Chudgar,
Jennifer Cuzzocreo,
Peter Calabresi,
Dzung Pham,
Daniel Reich,
Ciprian Crainiceanu