【速報】 Neuroinformatics 2016

イギリス・レディングで開催された Neuroinformatics 2016 に研究室から下記の3名が発表しました。

  • M1 萩原 里奈 Functional connectivity analysis of working memory during a mental arithmetic task
  • M1 吉武 沙規
  • M1 玉城 貴也 Region-of-interest estimation using convolutional neural network and long short-term memory for functional near-infrared spectroscopy data


【概要】
Neuroinformatics はNeuro Science + Informaticsであり、神経科学に関連する計算モデル、解析手法、およびそれらに関連する研究領域です。
Neuroinformatics 2016は、Neuroinformatics に関連した研究を議論する学会で、2016年はイギリス・レディングのレディング大学で開催れました。
レディングはイギリスの南、ロンドンの西の中都市で、最近はIT産業が集積しています。
毎年、夏8月に開催される野外ロック・フェスティバル レディング&リーズ・フェスティバル が有名で日本のロックフェスティバルもこのフェスを手本にしている。
Neuroinformatics の分野では、神経科学に関連する計算モデル、解析手法も重要であるが、それらのモデルや解析方法を広く共有するためには、それらを利用するデータを関係者で同じフォーマットで共有することが非常に重要である。この国際会議では、それらの内容についても議論が行われる珍しい会議である。
Neuroinformatics は参加者が500人程度のこじんまりとした学会であり、KeyNoteとPresentaion, Posterから構成され、2日間にわたって 最大2セッションでプログラムが構成されている。
【基調講演】
2日間で12セッションが行われ、各セッションに1件の基調講演があった。
【一般講演】
各セッションで数件の一般講演が行われた。
面白いのは、各セッション終了後に、講演を行った基調講演者、一般講演者が 全員参加し、パネルディスカッションを行うことである。
会場からも質問を受け付け、プレゼンターが答えるというおもしろいスタイルであった。
【ポスター】
会場を変えて、毎日 最終セッションでポスター講演が開催された。
会場は少し狭かったかな。
【おわりに】
 
—-
学会参加報告書

 報告者氏名 萩原里奈
発表論文タイトル Functional connectivity analysis of working memory during a mental arithmetic task
発表論文英タイトル Functional connectivity analysis of working memory during a mental arithmetic task
著者 萩原里奈,日和悟,廣安知之
主催 The Intermational Neuroinformatics Coordinating Facility
講演会名 Neuroinformatics2016
会場 Meadow Suite at Reading University
開催日程 2016/9/3-2016/9/4

 
 

  1. 講演会の詳細

2016/9/3-4に,イギリスのレディングにて開催されましたNeuroinformaticsに参加いたしました.この大会は,The Intermational Neuroinformatics Coordinating Facilityによって主催された大会で,データや知識ベースの神経システム,神経システムデータのためのツール,脳のモデル化に関する幅広いneuroinformaticsに携わる参加者が集まり,神経科学のツールの開発,神経科学データの処理方法について議論することを目的に開催されています.
本研究室からは他に廣安先生,日和先生,M1の吉武さん,玉城さんが参加しました.
 

  1. 研究発表
    • 発表概要

私は3日のポスターセッションに参加いたしました.発表の形式はポスター発表で,2時間自由に参加者の方と議論を行いました.
今回の発表は,「Functional connectivity analysis of working memory during a mental arithmetic task」と題して発表いたしました.以下に抄録を記載します.

Introduction: We spend daily life by using a storage system called working memory. Working memory (WM) has been defined as a system for the temporary holding and manipulation of the information (Baddeley and Hitch, 1974; Baddeley, 2000). It plays an important role in cognitive functions such as language recognition and reasoning ability. A limited memory buffer of retaining and processing the information is called as working memory capacity (WMC), and it is different between individuals. The individual differences in WMC affect a variety of cognitive activities. For example, poor WMC is said to be related to attention deficit hyperactivity disorder (ADHD), and working memory training to make WMC increased is used as treatment for ADHD patients (Klingberg et al., 2002). Working memory is used in the case of reading and solving mental arithmetic. Reading span test and N-back task are often used as an assessment of WMC. However, these paradigms are far from our daily life. Therefore, in this study, we adopted mental arithmetic task often used in everyday life. In a complex system such as working memory, each brain region does not always activate individually but often works cooperatively with other regions. Although the brain regions associated with working memory during mental arithmetic task had been revealed (Fehr et al., 2007), cooperative relationship among these regions have not been investigated enough. Therefore, in this study, we investigated the cooperative relationship among the brain regions during mental arithmetic task using a functional connectivity magnetic resonance imaging (fcMRI) study.Materials and Methods: Fourteen healthy adults (average age: 22.5 ± 1.5 years, 13 right-handed, 11 male) participated in this study. Participants performed the mental arithmetic task, which was consisted of the easy (non-working memory) and difficult (working memory) task, in the fMRI scanner. We calculated the correlation matrix and analyzed the functional connectivity. Acquired images were preprocessed by SPM8, and the activated regions were extracted and analyzed. A functional connectivity matrix of the individual data was calculated using Conn toolbox (Susan and Nieto-Castanon, 2012). Each image was partitioned into 116 regions using automatic anatomical labeling (AAL) atlas, and ROI-to-ROI connectivity was calculated for 116 regions. Moreover, graph theory metrics (degree and clustering coefficient) were calculated for each functional connectivity matrix using Brain Connectivity Toolbox (Mikail and Sporns, 2010). Degree is the number of edges connected to other nodes and indicates the centrality of a certain node (the brain region). Clustering coefficient indicates the degree to which nodes tend to cluster together. Moreover, the community of the brain network is extracted using Newman algorithm, which is a graph partitioning method for network analysis. These metrics allow us to quantitatively analyze the structure of brain network.Results and Discussions: Average correct answers of easy and difficult tasks were 30.7 ± 4.9 and 3.88 ± 1.17%, respectively, and it was confirmed that the answers of easy task were significantly higher than those of difficult task [t (13) = 22.5, p < 0.01). By group analysis, we performed a paired t-test to examine the difference of activation between easy task and difficult task. Supplementary motor area and middle temporal gyrus were extracted as the regions whose activations were significantly higher for difficult task than for easy task. Significant activated regions during the difficult task were both cuneus and left precuneus. Since these regions are associated with working memory (Tomasi et al., 2006; Neta and Whalen, 2011; Sala-Llonch et al., 2012), they activated during the difficult task. Moreover, degree measures of these regions were significantly higher for difficult task than for easy task (paired t-test, p < 0.05). This suggested that these regions worked cooperatively with other regions during the difficult task compared with during the easy task. On the other hand, the clustering coefficient is higher for difficult task than for easy task for all regions (p < 0.05). This indicated that brain regions tended to form a cluster during the difficult task. Furthermore, 116 brain regions were partitioned into four communities, the frontal lobe, the parietal lobe, the temporal lobe and the occipital lobe in easy task. In difficult task, the frontal lobe and the parietal lobe were categorized into the same group, so that three communities were identified. These results suggested that the frontal lobe and the parietal lobe needed to work together during working memory task. In addition, observation of the regions connected from the both cuneus revealed that there were connections from the both cuneus to the frontal lobe (left middle frontal gyrus) and the parietal lobe (left inferior parietal lobule, left paracentral lobule) during difficult task. The frontal–parietal network is associated with visual attention, and controls the occipital visual cortex to selectively process only the required visual information (Nobre et al., 1997). Since our results found connections between the both cuneus and the frontal and parietal lobes, it suggested that the cuneus (the occipital visual cortex) was controlled by the frontal lobe and the parietal lobe.
 

 

  • 質疑応答

今回の講演発表では,以下のような質疑を受けました.
 
・質問内容1
質問者の氏名を控え損ねてしまいました.こちらの質問は相関行列の閾値はいくつで,エッジの値はバイナリであるかというものでした.閾値は0とし,negative connectivityは検討せず,positive connectivityのみみており,エッジは相関係数の重みがあると回答致しました.
 
・質問内容2
質問者の氏名を控え損ねてしまいました.こちらの質問はコネクティビティは実験設計全体の時系列データを用いているのか,あるいはタスクの時系列データのみ用いているのかというものでした.こちらの質問に対する回答ですが,各タスクの時系列データを抽出し連結していると回答致しました.
 
・質問内容3
 質問者の氏名を控え損ねてしまいました.こちらの質問は,ネットワーク特徴量はStrengthのみを用いているかというものでした.この質問に対して,今回はStrengthのみの検討であり,今後Clustering coefficientやModularityなどの特徴量も検討していきたいと回答致しました.
 
・質問内容4
株式会社ドワンゴ所属の山川さんからの質問です.こちらの質問は,コネクティビティの違いはStrengthを検討する前に,直接的にコネクティビティの違いを検討していないのかというものでした.また,今回の結果のコネクティビティの違いは先行研究で示されているかという質問をいただきました.今回は,直接的なコネクティビティの比較は目視のみで行っているため,今後検討したいと思うと回答致しました.また,抽出された脳領域の機能に関してのみ先行研究で調べが,コネクティビティに関する先行研究は調べられていないと回答致しました.今後検討したいと思います.
 

  • 感想

2度目の学会発表でしたが,今回は初めての国際学会であり,初めてのポスター発表であったため,とても緊張しました.日和先生のご指導のおかげで,インパクトのあるポスターを作成することができたため,多くの方にポスターをみていただくことができました. どの方もとても優しく,たどたどしい英語でしたが,なんとか自分の研究を英語で伝えることができたのではないかと思います.コネクティビティの研究や,グラフ理論を用いた研究も多くあり,自分の知識が不十分であることを感じました.
 

  1. 聴講

今回の講演会では,下記の2件の発表を聴講しました.
 

発表タイトル       : Whole brain fMRI activity at a high temporal resolution: A novel analytic framework著者                  : Niels Janssenセッション名       : Track C – Neuroimaging I
Abstruct            :
We have developed a new framework for the analysis of functional Magnetic Resonance Imaging (fMRI) data. Whereas current analytic techniques primarily yield static, time-invariant maps of fMRI activity (Smith et al., 2004), our new technique yields dynamic, time-variant videos of whole-brain fMRI activity. The new framework relies on a fundamentally different method of fMRI BOLD signal extraction. Specifically, instead of the standard volume-based signal extraction, the new method extracts the fMRI BOLD signal based on the veridical MRI slice acquisition times. This yields an fMRI signal that is more temporally accurate (Sladky et al., 2011). In addition, we improved the temporal resolution by presenting each slice to a different point in the progression of the BOLD signal [see also Price et al. (1999)]. The fMRI BOLD signal is then extracted using non-standard statistical modeling techniques. Specifically, the fMRI data are first broken up into epochs that are time-locked to the onset of a stimulus. Next, in line with techniques used in EEG (Janssen et al., 2014), statistical models are run at each time-point in the epoch. As the baseline, we used the fMRI signal intensity values available at time-point 0. For this particular choice of baseline, modeling involves extracting the fMRI BOLD signal across time points in the epoch. The number of available timepoints in the epoch (and therefore the temporal resolution) is scalable, up to a maximum that is determined by the rate at which MRI slices are acquired (typically on the order of tens of milliseconds). In order to account for the full complexity of the statistical model, we used Linear Mixed Effect modeling (Pinheiro and Bates, 2000). Our method yields an fMRI signal for every voxel in the brain that is more temporally accurate and of a much higher temporal resolution that is available in current frameworks.
The data manipulation in the new framework relies on functions written as part of the neuro-imaging data analysis package FSL (Smith et al., 2004) and various Python scripts of which the NiBabel package for reading neuro-imaging data forms an indispensable part (http://nipy.org/nibabel/). Statistical modeling of first order individual participant data relied on the data.table and lme4 packages available in the software R (Douglas et al., 2015). Higher order modeling was performed with the randomise function of FSL (Winkler et al., 2014). A key characteristic of the current approach is that it does not rely on data averaging but uses all data points from all epochs in an experiment to model the signal. Advantages of using this pipeline are that statistical modeling of first-order fMRI data is greatly simplified and handled by R. Disadvantages are the slow speed of R, and the large file sizes due to the long data table format requirements imposed by R. We will illustrate the new technique in the context of fMRI data collected during a visual object naming experiment. We will use these data to explore the spatio-temporal dynamics of the whole-brain fMRI BOLD signal at 390 ms temporal resolution, focusing on task-based functional connectivity. Our new framework can be easily applied to data collected with other types of tasks and provides a novel opportunity to gain insight into the spatio-temporal dynamics of fMRI activity during cognitive tasks.
 
Acknowledgment
This work was supported by The Spanish Ministry of Economy and Competitiveness (RYC2011-08433 and PSI2013-46334 to NJ)

fMRIの撮像時間に関する問題に関する発表でした.1ボリュームでスライス毎に時間が違うことと1ボリューム毎の時間分解能が低いことの両方に関してアプローチするもの,Slice-Based手法により同じタイミングのスライスで1ボリュームを撮像するものでした.実験で用いているMRIの時間分解は高くないため,解析の際にはスライスタイミングについても考える必要があると再確認させられました.
 

発表タイトル       : Measuring complex brain networks structure著者                  : Ester Bonmati, Anton Bardera, Imma Boadaセッション名       : Track E – Informatics III: Visualization
Abstruct            :
Introduction: The human brain has roughly one hundred billion neurons forming a network with trillions of intra-connections. The mapping of structure and functionality of brain networks is therefore an important challenge in understanding the functioning. Connectivity matrices are used to represent brain networks, also called connectome (Hagmann, 2005; Sporns et al., 2005), as a graph (Hagmann et al., 2007, 2010; Sporns, 2013), where nodes correspond to brain regions and edges to structural or functional connections (Bullmore and Bassett, 2011; Sporns, 2011; Wu et al., 2013).
Different measures have been applied to describe topological features of brain networks (Stam and Reijneveld, 2007; Rubinov and Sporns, 2010; Kaiser, 2011). For instance, the independence of large areas, denoted as integration, has been studied by the path length measure, the characteristic path length (Watts and Strogatz, 1998), or the global efficiency (Latora and Marchiori, 2001). Independence of small subsets, defined as segregation, can be analyzed by the clustering coefficient (Watts and Strogatz, 1998), the transitivity (Newman, 2003a), or the modularity (Newman, 2003b). The importance of individual nodes can be defined with centrality measures such as the degree (Bullmore and Sporns, 2009). A good summary of the measures can be found in Rubinov and Sporns (2010).
In this work, we present a global and two local measures, based on the mutual information measure, to quantify brain networks structure.
Materials and Methods: Materials: Synthetic model networks were created using the Brain Connectivity Toolbox (BCT) (Rubinov and Sporns, 2010). Random, lattice, ring lattice, and small-world model networks with 128 and 256 nodes with edges ranging from 128 to 8192 with a step of 128 edges were used. Additionally, networks with nodes ranging from 32 to 512 with a step of 32 and a density of 0.4 were also created.
As human structural networks, we used the normalized connection matrices created from MRI tractography described in Cammoun et al. (2012). As human functional networks, we used the HCP 500-PTN functional dataset (Van Essen et al., 2012; Glasser et al., 2013; Hodge et al., 2015).
All networks were weighted and non-directed.
Method: In the proposed approach, brain networks are modelled as a Markov process where neuronal impulses randomly walk from one node to another node. This new interpretation provides a solid theoretical framework from which we derive a global (i.e., a single value for the whole network) and two local (i.e., a value for each node) measures based on mutual information.
Mutual information (MI) measures the shared information between two random variables. From our Markov process-based brain model, we propose as a global connectivity measure the mutual information between two consecutive states of the process. Mutual information can also be seen as the difference between the uncertainty of the states without any knowledge and the uncertainty of the states when the past is known (or information gained when the previous node is known). The higher the MI, the less random the connections. Thus, mutual information can be used to quantify the overall brain structure.
The mutual information can be decomposed in order to characterize the degree of informativeness of each state. When applied to the connectome, since each state corresponds to an anatomical or functional region, this measure can be seen as the contribution of each node to the whole graph structure. In this work, we propose two local measures. On the one hand, we use the mutual surprise (I1) (DeWeese and Meister, 1999), that expresses how “surprising” are the connections of a node. Nodes that are connected with more likely nodes will lead to low values of mutual surprise, while those with very specific connections or connected with few unlikely nodes will have high mutual surprise. On the other hand, we use the mutual predictability (I2) (DeWeese and Meister, 1999), that expresses the uncertainty of a node taking into account the mean connectivity of all the network. I2 measures the capacity of prediction for a given brain region.
Results and Discussion: Using model networks with different number of edges, an optimum point was found for lattice and ring lattice networks when increasing the density. This is due to the fact that for low densities, there are regions not connected, thus, the overall mutual information is low. This fact may help to find a minimum number of fibers needed to study brain networks for a given brain parcellation. Overall, higher values were obtained for lattice and ring lattice models, showing a clear evidence of more organized networks compared with random and small-world networks. When the number of edges was increased, the mutual information tended to decrease, since the higher number of connections, the lower correlation between consecutive states. Preserving the density, the mutual information was not very sensitive to random and small-world networks, since the structure is similar. Higher values were obtained for ring lattice networks when comparing with lattice networks, since in lattice networks two nodes are not connected and have a less structured network. Using anatomical and functional connectomes at different scales, a similar behavior was observed for all patients.
Local measures were evaluated using the human connectomes. The mutual surprise highlighted regions connected to regions not highly connected, such as the right hemisphere transverse temporal. Low values were obtained for regions connected to highly connected regions such as the left hemisphere thalamus proper. The mutual predictability associated regions with a low number of connections and high weights with a high predictability, such us the right hemisphere temporal pole. Low values were obtained in regions with more uncertainly in predicting the next node, such as the right hemisphere putamen.
All measures were consistent for structural and functional human networks.
Conclusion: In this work, new measures to quantify structure of complex brain networks are proposed. Brain connectivity graphs are interpreted as a stochastic process where neural impulses are modeled as a random walk. This interpretation provides a solid theoretical framework from which different measures based on the mutual information measure have been applied.
The measures have been tested on synthetic model networks and structural and functional human networks at different scales. Results show that the mutual information is able to quantify the structure of different model networks. The mutual surprise, allows the identification of nodes whose neighbors have a high connectivity taking into account all connections. The mutual predictability shows that regions with a high clustering tend to be more predictable.
Acknowledgments
This work was supported by the Spanish Government (Grant No. TIN2013-47276-
C6-1-R) and by the Catalan Government (Grant No. 2014-SGR-1232). Data were provided, in part, by the Human Connectome Project, WU-Minn Consortium (Principal Investigators: David Van Essen and Kamil Ugurbil; 1U54MH091657) funded by the 16 NIH Institutes and Centers that support the NIH Blueprint for Neuroscience Research; and by the McDonnell Center for Systems Neuroscience at Washington University.

本発表は,構造的コネクティビティをグラフ理論を用いて解析するものでした.私は機能的コネクティビティをグラフ理論によって解析しているので,関連する発表でしたが,マルコフ連鎖を用いることで脳の状態を検討していました.グラフ理論だけでなく他の解析方法を用いることで,より深い研究を行っていきたいと感じました.
 
参考文献

  • Neuroinformatics 2016,

http://neuroinformatics2016.org/
 
学会参加報告書

報告者氏名 玉城貴也
発表論文タイトル Region-of-interest estimation using convolutional neural network and long short-term memory for functional near-infrared spectroscopy data
発表論文英タイトル Region-of-interest estimation using convolutional neural
network and long short-term memory for functional
near-infrared spectroscopy data
著者 玉城貴也,日和悟,蜂須賀啓介,奥野英一,廣安知之
主催 The Intermational Neuroinformatics Coordinating Facility
講演会名 Neuroinformatics2016
会場 Meadow Suite at Reading University
開催日程 2016/9/3-2016/9/4

 
 

  1. 講演会の詳細

2016/9/3-4に,イギリスのレディングにて開催されましたNeuroinformaticsに参加いたしました.この大会は,The Intermational Neuroinformatics Coordinating Facilityによって主催された大会で,データや知識ベースの神経システム,神経システムデータのためのツール,脳のモデル化に関する幅広いneuroinformaticsに携わる参加者が集まり,神経科学のツールの開発,神経科学データの処理方法について議論することを目的に開催されています.
本研究室からは他に廣安先生,日和先生,M1の吉武さん,萩原さんが参加しました.
 

  1. 研究発表
    • 発表概要

私は3日のポスターセッションに参加いたしました.発表の形式はポスター発表で,2時間自由に参加者の方と議論を行いました.
今回の発表は,「Region-of-interest estimation using convolutional neural network and long short-term memory for functional near-infrared spectroscopy data」と題して発表いたしました.以下に抄録を記載します.

Introduction: In recent years, functional near-infrared spectroscopy (fNIRS) has attracted attention as a noninvasive functional neuroimaging technology. fNIRS visualizes brain activity by measuring the hemodynamic responses of oxy- and deoxyhemoglobin (Hb) associated with neural behavior. fNIRS allows us to identify a cortical activation or brain regions associated with a given stimulus by analyzing the time courses of oxy- and deoxy-Hb. However, since fNIRS signals often contain noises (e.g., motion-related artifacts and psychological noises including heartbeat), it is not easy to extract meaningful brain activation. Furthermore, comparison of the raw fNIRS data between subjects should not be done because fNIRS detects only the relative change in oxy- or deoxy-Hb. To perform a group analysis, baseline calibration is also needed. However, there is no established way for preprocessing of fNIRS signals. The purpose of this study was to estimate the brain regions associated with a given task or stimulus by automatically extracting features of cortical activities from the fNIRS data. We proposed a novel feature extraction method for the fNIRS signals whose temporal and spatial characteristics were considered.
Method: Deep learning methods have been mostly used for classification of multi-dimensional data, however, in this study, we focused on another aspect of the deep learning methodology regarding determination of region-of-interest (ROI) associated with a given task or stimuli. In our proposed approach, a group classifier is constructed from all subject fNIRS data using supervised learning. The group classifier is constructed for all channels of a fNIRS measurement system, and a group label is supervised during each learning process. After the learning is completed, the classification accuracy using only a single channelis compared among all channels, and the channel whose classification accuracy hasbetter performance is extracted as the critical ROI for group classification.Moreover, we proposed a new deep learning algorithm which is a fusion of two algorithms,convolutional neural network (CNN) (LeCun, 1998) and long short-termmemory (LSTM) (Hochreiter and Schmidhuber, 1997). Although both algorithmscan automatically perform feature extraction, CNN preserves the spatial informationon input data during learning, and LSTM stores the temporal information. Takingadvantage of these two algorithms, our proposed algorithm basically consisted of fivelayers, input, convolution, LSTM, pooling, and output layer. We can identify the ROIbecause neuron units of the input layer are associated with the fNIRS probes placedon participants’ head.
Experiments: To examine the effectiveness of our approach, we tried to extract theROIs related to working memory. Cerebral blood flow during N-back (N = 2, 3) task,which was often used to assess the working memory, was measured using fNIRS. 30healthy male subjects (average age: 23.3 ± 1.5 years, right-handed) and 5 healthy femalesubjects (average age: 21.7 ± 0.52 years, right-handed) participated in the experiment.The fNIRS probes were placed according to the International 10–20 system. Usingthe fNIRS data obtained, our classifier is trained to classify the input data as either“2-back” or “3-back.”
Results and Discussion: The average percentage of correct answers in 2-back (low degree of difficulty) and 3-back (high degree of difficulty) tasks were 90.2 ± 8.98 and 84.3  ±  8.87%, respectively. It was shown to be significantly different by Wilcoxon signed-rank test (p < 0.01). Using this fNIRS data, our classifier achieved the classification accuracy of 91.4 ± 1.49%. Moreover, with a comparison of single-channel classification accuracy for all channels, we successfully extracted left dorsolateral prefrontal cortex (DLPFC) and anterior prefrontal cortex (APFC) as task-related ROIs. DLPFC is activated in a number of working memory task and cognitive task, and is also known to play a key role in cognitive control and adaptation of a strategy to improve the task performance. In particular, the left DLPFC is said to be activated in verbal working memory task (Smith and Jonides, 1997).APFC is the area where highly abstract information is processed. It has also been reported that APFC and DLPFC had activated in a dual-task situation (Narender and Owe, 2004). Furthermore, activation of DLPFC and APFC is associated with the difficulties of N-back problems (Owen, 2005; Katrin et al., 2014). These observations suggest that the ROIs estimated by the proposed method are reasonable. Consequently, our proposed method has been shown to be useful for a brain function analysis of fNIRS data.

 

  • 質疑応答

今回の講演発表では,以下のような質疑を受けました.
 
・質問内容1
質問者の氏名を控え損ねてしまいました.こちらの質問はDeep Learningというと中身はブラックボックスとされているが,どのようにしてROIを推定するのかという質問でした.こちらの質問に対する回答ですが,DeepLearningを始めとする機械学習アルゴリズムには入力と出力をつなぐパラメータが存在し,そのパラメータを見て,入力までたどることで,どの部位が識別に重要であるかを判断できると回答致しました.
・質問内容2
ドワンゴ人工知能研究所所長の山川さんからの質問です.こちらの質問は,なぜ空間性と時系列性を分けて行う必要があるのかという質問でした.こちらの質問に対する回答ですが,空間性と時系列性を同時に行うと,どこに識別に必要な特徴量があったのかを特定することが困難であり,構造的に特定しやすいために,分けて行ったと回答致しました.
 

  • 感想

今回は初めての国際学会であり,英語で自分の研究内容が伝わるのかという不安がありました.
しかし,日和先生のご指導のおかげで,インパクトのあるポスターを作成することができ,多くの方の目に止まり,その際に自分の研究の中で一番伝えたいことを重点に説明することで多くの方に理解してもらうことができました.そして,学会参加することで,自分の発表に対する貴重なご意見や質問をいただき,大変勉強になりました.
 

  1. 聴講

今回の講演会では,下記の2件の発表を聴講しました.
 

発表タイトル       : Measuring complex brain networks structure
著者                  : Ester Bonmati, Anton Bardera, Imma Boada
セッション名       : Track E – Informatics III: Visualization
Abstruct            :
Introduction: The human brain has roughly one hundred billion neurons forming a network with trillions of intra-connections. The mapping of structure and functionality of brain networks is therefore an important challenge in understanding the functioning. Connectivity matrices are used to represent brain networks, also called connectome (Hagmann, 2005; Sporns et al., 2005), as a graph (Hagmann et al., 2007, 2010; Sporns, 2013), where nodes correspond to brain regions and edges to structural or functional connections (Bullmore and Bassett, 2011; Sporns, 2011; Wu et al., 2013).
Different measures have been applied to describe topological features of brain networks (Stam and Reijneveld, 2007; Rubinov and Sporns, 2010; Kaiser, 2011). For instance, the independence of large areas, denoted as integration, has been studied by the path length measure, the characteristic path length (Watts and Strogatz, 1998), or the global efficiency (Latora and Marchiori, 2001). Independence of small subsets, defined as segregation, can be analyzed by the clustering coefficient (Watts and Strogatz, 1998), the transitivity (Newman, 2003a), or the modularity (Newman, 2003b). The importance of individual nodes can be defined with centrality measures such as the degree (Bullmore and Sporns, 2009). A good summary of the measures can be found in Rubinov and Sporns (2010).
In this work, we present a global and two local measures, based on the mutual information measure, to quantify brain networks structure.
Materials and Methods: Materials: Synthetic model networks were created using the Brain Connectivity Toolbox (BCT) (Rubinov and Sporns, 2010). Random, lattice, ring lattice, and small-world model networks with 128 and 256 nodes with edges ranging from 128 to 8192 with a step of 128 edges were used. Additionally, networks with nodes ranging from 32 to 512 with a step of 32 and a density of 0.4 were also created.
As human structural networks, we used the normalized connection matrices created from MRI tractography described in Cammoun et al. (2012). As human functional networks, we used the HCP 500-PTN functional dataset (Van Essen et al., 2012; Glasser et al., 2013; Hodge et al., 2015).
All networks were weighted and non-directed.
Method: In the proposed approach, brain networks are modelled as a Markov process where neuronal impulses randomly walk from one node to another node. This new interpretation provides a solid theoretical framework from which we derive a global (i.e., a single value for the whole network) and two local (i.e., a value for each node) measures based on mutual information.
Mutual information (MI) measures the shared information between two random variables. From our Markov process-based brain model, we propose as a global connectivity measure the mutual information between two consecutive states of the process. Mutual information can also be seen as the difference between the uncertainty of the states without any knowledge and the uncertainty of the states when the past is known (or information gained when the previous node is known). The higher the MI, the less random the connections. Thus, mutual information can be used to quantify the overall brain structure.
The mutual information can be decomposed in order to characterize the degree of informativeness of each state. When applied to the connectome, since each state corresponds to an anatomical or functional region, this measure can be seen as the contribution of each node to the whole graph structure. In this work, we propose two local measures. On the one hand, we use the mutual surprise (I1) (DeWeese and Meister, 1999), that expresses how “surprising” are the connections of a node. Nodes that are connected with more likely nodes will lead to low values of mutual surprise, while those with very specific connections or connected with few unlikely nodes will have high mutual surprise. On the other hand, we use the mutual predictability (I2) (DeWeese and Meister, 1999), that expresses the uncertainty of a node taking into account the mean connectivity of all the network. I2 measures the capacity of prediction for a given brain region.
Results and Discussion: Using model networks with different number of edges, an optimum point was found for lattice and ring lattice networks when increasing the density. This is due to the fact that for low densities, there are regions not connected, thus, the overall mutual information is low. This fact may help to find a minimum number of fibers needed to study brain networks for a given brain parcellation. Overall, higher values were obtained for lattice and ring lattice models, showing a clear evidence of more organized networks compared with random and small-world networks. When the number of edges was increased, the mutual information tended to decrease, since the higher number of connections, the lower correlation between consecutive states. Preserving the density, the mutual information was not very sensitive to random and small-world networks, since the structure is similar. Higher values were obtained for ring lattice networks when comparing with lattice networks, since in lattice networks two nodes are not connected and have a less structured network. Using anatomical and functional connectomes at different scales, a similar behavior was observed for all patients.
Local measures were evaluated using the human connectomes. The mutual surprise highlighted regions connected to regions not highly connected, such as the right hemisphere transverse temporal. Low values were obtained for regions connected to highly connected regions such as the left hemisphere thalamus proper. The mutual predictability associated regions with a low number of connections and high weights with a high predictability, such us the right hemisphere temporal pole. Low values were obtained in regions with more uncertainly in predicting the next node, such as the right hemisphere putamen.
All measures were consistent for structural and functional human networks.
Conclusion: In this work, new measures to quantify structure of complex brain networks are proposed. Brain connectivity graphs are interpreted as a stochastic process where neural impulses are modeled as a random walk. This interpretation provides a solid theoretical framework from which different measures based on the mutual information measure have been applied.
The measures have been tested on synthetic model networks and structural and functional human networks at different scales. Results show that the mutual information is able to quantify the structure of different model networks. The mutual surprise, allows the identification of nodes whose neighbors have a high connectivity taking into account all connections. The mutual predictability shows that regions with a high clustering tend to be more predictable.
Acknowledgments
This work was supported by the Spanish Government (Grant No. TIN2013-47276-
C6-1-R) and by the Catalan Government (Grant No. 2014-SGR-1232). Data were provided, in part, by the Human Connectome Project, WU-Minn Consortium (Principal Investigators: David Van Essen and Kamil Ugurbil; 1U54MH091657) funded by the 16 NIH Institutes and Centers that support the NIH Blueprint for Neuroscience Research; and by the McDonnell Center for Systems Neuroscience at Washington University.

本発表は,脳の状態の遷移をマルコフ過程としてモデル化し,構造的コネクティビティをグラフ理論を用いて解析するものでした.脳の状態を解析する上でマルコフ連鎖も検討する必要があると感じました.
 

発表タイトル       : Functional connectivity of resting state as a biomarker for working memory performance
著者                  : Hüden Nese, Can Soylu, Pınar Adanalı, Metahan Irak, Ata Akın
セッション名       : Poster
Abstruct            :
Graph theory is a powerful tool to investigate the brain as a complex system. Several studies have worked on the relation between individual difference in cognitive ability and the network organization of the brain (Li et al., 2009; van den Heuvel et al., 2009). However, this relation is still poorly understood.
In this study, we investigated the relation between resting state functional connectivity and performance in a working memory task. Spontaneous EEG of 15 subjects were recorded for 2 min with closed eyes using 32-channel system. Then participants per- formed a verbal n-back task. The correlation coefficient between channel pairs during the rest period was used as a measure of functional connectivity. Connectivity matrix was thresholded according the predefined connection density. Global efficiency, local efficiency, and modularity were computed from the processed connectivity matrix, normalized, and statistically analyzed.
We showed that there is a relation between functional integration and cognitive per- formance. Global efficiency is significantly correlated with performance especially in beta (r = 0.52, p = 0.045) and gamma (r = 0.58, p = 0.023) bands. Besides, there is also significant correlation between modularity and performance in beta band (r = 0.55, p = 0.33). Our results are mainly compatible with previously published studies where they reported that global efficiency and modularity are the predictors of cognitive performance regardless of the cognitive task domain (Langer et al., 2012; Stevens et al., 2012; Alavash et al., 2015). However, there are conflicting results on the local efficiency depending on the cognitive task (van den Heuvel et al., 2009; Langer et al., 2012).

Working Memoryのネットワーク結合に関する発表でした.認知課題時の個人差の関係性を解析により,調査するというものでした.今まで,個人差については検討をしていなかったので,個人で脳機能状態がどのように変化するのか今後検討する必要があると感じました.
 
参考文献

  • Neuroinformatics 2016,

http://neuroinformatics2016.org/
 
学会参加報告書

 
報告者氏名
 
吉武 沙規
発表論文タイトル Adaptive HRF and BF approaches to fNIRS activation analysis
発表論文英タイトル Adaptive HRF and BF approaches to fNIRS activation analysis
著者 吉武 沙規
主催 医療情報システム研究室
講演会名 Neuroinformatics2016
会場 University of Reading
開催日程 2016/09/03-2016/09/04

 
 

  1. 講演会の詳細

2016/09/03から2016/09/04にかけて,Reading Universityにて開催されましたNeuroinformatics2016に参加いたしました.他にこの学会は,incfによって主催された学会で,この学会は,結果の共有と解析方法の提供を目的として開催されました.またこの学会内では,コンピュータシステムや画像化,ネットワークのモデリングがセッション内で発表されました.
本研究室から他に,廣安先生,日和先生,M1の萩原さん,玉城さんが参加しました.
 

  1. 研究発表
    • 発表概要

私は3日の17:30からのPOSTER AND DEMO RECEPTIONに参加いたしました.ポスター前での議論が行われました.
今回の発表は,fNIRSによる計測データの解析方法についての発表を行いました.以下に抄録を記載致します.

Functional Near Infrared Spectroscopy is one of the measurement methods for elucidation cerebral function. Measuring cerebral bloodflow change by using fNIRS, brain activation can be judged. GLM is one of the judging methods of brain activation using cerebral bloodflow change. In GLM, brain activation is judged by regression analysis of hemodynamic model and measurement data[1][2]. This hemodynamic model is created by convolution of hemodynamic response function(HRF)[3] and rectangular function based on the experimental design. Rectangular function and HRF used in thin method is the same, regardless subjects, measurement region and experimental design. But there is a possibility HRF varies depend on brain region and tasks and within subjects[3][4]. Therefore, each HRF is not always the same as general HRF. The method using hemodynamic model made with same HRF and rectangular function can’t analyze along measurement data. So, it is considered that this method is unconvincing in judging brain activation. Therefore, accurate judging method is needed. Rectangular function containing variation information of cerebral activation can be determined by optimizing the function based on measurement data. Also, measurement data can be expressed exactly by optimizing HRF containing timing information based on rectangular function and measurement data. In this way, brain activation is judged exactly. Hemodynamic model matched with fNIRS data was made by optimizing HRF after optimizing rectangular function. HRF parameters are the first peak delay, the undershoot delay and amplitude ratio between the first peak and the undershoot. In optimizing rectangular function, the size of function was determined as the amount of change after 5 seconds which is maximum arrival time. Regression analysis performs on measurement data and hemodynamic model made with rectangular function and HRF. Index of optimization is t value of regression coefficient. HRF parameters are determined when t value reaches maximum. Using 3 parameters of HRF, cerebral function of each subjects and cerebral region was examined.
 

 

  • 質疑応答

今回の講演発表では,以下のような質疑を受けました.
 
・質問内容1
使用機器であるNIRSについての質問をいただきました.したがってNIRSによって計測しているデータが実際に何を計測しどのような内容を表しているのかを説明しました,また,計測原理について近赤外光を使用していることなどを説明しました.
 
・質問内容2
メインの結果が何を示しているのかを説明してほしいという質問がありました.したがって私は,この研究のコンセプトから説明を行いました.コンセプトと比較しながら実際に出た結果が実行課題の刺激の大きさを示すことができ,時系列の変化によってどのように変わっているのか,またHRFの最適化によって刺激に対する血流反応の速さの検討が可能になることを説明しました.
 
・質問内容3
Cross-validationを使用しているのかを質問されました.したがって私はこの研究では回帰分析を使用しているので,使用していないと返答しました.
 
・質問内容4
課題であるn-backタスクを知らない方が多かったので,課題の説明と難易度によりどう違うのかを説明しました.
・質問内容5
今回解析しているデータはOxy-Hbのみで,Deoxy-Hbは使用していないのかと聞かれたので
Oxy-Hbのみ使用していると返答しました.理由としてDeoxy-HbでのHRFについてあまりわかっていないことを説明しました.
 

  • 感想

発表前は自分の英語がどこまで通じるのか,きちんと返答できるのかと不安が大きかったですが,実際に発表した際には,聞き返すと分かりやすくゆっくりと質問しなおしてくださったりして, 受け答えができていたと思います.また,自分が想像していたよりもNIRSやHRFについてご存じない方が多かったので,説明資料を作成していくべきだったと思いました.
 

  1. 聴講

今回の講演会では,下記の2件の発表を聴講しました.
 

発表タイトル       : Whole brain fMRI activity at a high temporal resolution: A novel analytic framework
著者                  :  Niels Janssen
セッション名       : Neuroimaging I
Abstruct            : We have developed a new framework for the analysis of functional Magnetic Resonance Imaging (fMRI) data. Whereas current analytic techniques primarily yield static, time-invariant maps of fMRI activity (Smith et al., 2004), our new technique yields dynamic, time-variant videos of whole-brain fMRI activity. The new framework relies on a fundamentally different method of fMRI BOLD signal extraction. Specifically, instead of the standard volume-based signal extraction, the new method extracts the fMRI BOLD signal based on the veridical MRI slice acquisition times. This yields an fMRI signal that is more temporally accurate (Sladky et al., 2011). In addition, we improved the temporal resolution by presenting each slice to a different point in the progression of the BOLD signal [see also Price et al. (1999)]. The fMRI BOLD signal is then extracted using non-standard statistical modeling techniques. Specifically, the fMRI data are first broken up into epochs that are time-locked to the onset of a stimulus. Next, in line with techniques used in EEG (Janssen et al., 2014), statistical models are run at each time-point in the epoch. As the baseline, we used the fMRI signal intensity values available at time-point 0. For this particular choice of baseline, modeling involves extracting the fMRI BOLD signal across time points in the epoch. The number of available timepoints in the epoch (and therefore the temporal resolution) is scalable, up to a maximum that is determined by the rate at which MRI slices are acquired (typically on the order of tens of milliseconds). In order to account for the full complexity of the statistical model, we used Linear Mixed Effect modeling (Pinheiro and Bates, 2000). Our method yields an fMRI signal for every voxel in the brain that is more temporally accurate and of a much higher temporal resolution that is available in current frameworks.
The data manipulation in the new framework relies on functions written as part of the neuro-imaging data analysis package FSL (Smith et al., 2004) and various Python scripts of which the NiBabel package for reading neuro-imaging data forms an indispensable part (http://nipy.org/nibabel/). Statistical modeling of first order individual participant data relied on the data.table and lme4 packages available in the software R (Douglas et al., 2015). Higher order modeling was performed with the randomise function of FSL (Winkler et al., 2014). A key characteristic of the current approach is that it does not rely on data averaging but uses all data points from all epochs in an experiment to model the signal. Advantages of using this pipeline are that statistical modeling of first-order fMRI data is greatly simplified and handled by R. Disadvantages are the slow speed of R, and the large file sizes due to the long data table format requirements imposed by R.
We will illustrate the new technique in the context of fMRI data collected during a visual object naming experiment. We will use these data to explore the spatio-temporal dynamics of the whole-brain fMRI BOLD signal at 390 ms temporal resolution, focusing on task-based functional connectivity. Our new framework can be easily applied to data collected with other types of tasks and provides a novel opportunity to gain insight into the spatio-temporal dynamics of fMRI activity during cognitive tasks.

この発表ではMRIデータの解析手法を提案していました.刺激を与えた時のスライスを用いてボリュームの再構成をしていました.この方法を使用すると刺激付近の脳状態を観察することが可能になっていました.しかし,この方法では刺激による脳活動の変化のみを見られるとは限らないので,検討を行う必要があると感じました.

発表タイトル       :Mechanisms underlying different onset patterns of focal seizures
著者                  : Yujiang Wang
セッション名       : Brain Disorders I
Abstruct            : Focal seizures typically begin with an electrographic onset pattern that is highly stereotyped in individual patients. Qualitative classifications of these onset patterns describe two frequently occurring waveforms – low amplitude fast oscillations (LAF), or high amplitude spikes (HAS). Interestingly, only the former of the patterns is associated with a good surgical outcome. Given the importance of this clinical distinction, we therefore explored whether these two patterns arise from fundamentally different  spatio-temporal dynamics.
We used a previously established computational model of neocortical tissue, and validated it as an adequate model using clinical recordings of focal seizures. Using this model we then investigated the possible mechanisms underlying the different focal seizure onset patterns.
We show that the two patterns are associated with different mechanisms at the spatial scale of a single ECoG electrode. The LAF onset is initiated by independent patches of localised activity, which slowly invade the surrounding tissue and coalesce over time (see Figure 1A). In contrast, the HAS onset is a global, systemic transition to a coexisting seizure state triggered by a local event (see Figure 1B). We find that such a global transition is enabled by an increase in excitability of the surrounding tissue, essentially creating a seizure supporting surrounding. In our simulations, the difference in surrounding tissue excitability also offers a simple explanation of the clinically observed difference in surgical outcomes. Finally, we demonstrate how changes in tissue excitability could be elucidated in principle using active stimulation.
We conclude that the excitability of the tissue surrounding the seizure core plays a determining role in the seizure onset pattern, as well as in the surgical outcome.

この発表では,てんかんの発作発生時の脳活動をEEGで計測し研究していました.また,その時のニューロン変化を見ていました.てんかんの発生には特定の細胞が起因しているのではないかと述べていました.てんかんは多くの方がかかっている病気でもあるので,研究が進み少しでもQOLが高くなるといいと思っているので,このような研究をしている発表者が他にもいて興味を持って聞くことができました.
 
参考文献
(1) S. Tsujimoto, T. Yamamoto, H. Kawaguchi, H. Koizumi and T. Sawaguchi, “Prefrontal cortical activation associated with working memory in adults and preschool
children: an event-related optical topography study,” Neuroimage, vol. 1, no. 21,

  1. 283–290, 2004.

 
(2) M. Hofmann, M. Herrmann, I. Dan, H. Obrig, M. Conrad, L. Kuchinke, A. Jacobs
and A. Fallgatter, “Differential activation of frontal and parietal regions during visual
word recognition: an optical topography study,” Neuroimage, vol. 3, no. 40, pp.
1340–1349, 2008.
 
(3) T. Sano, D. Tsuzuki, I. Dan, H. Dan, H. Yokota, K. Oguro and E. Watanabe, “Adaptive hemodynamic response function to optimize differential temporal information
of hemoglobin signals in functional near-infrared spectroscopy,” Complex Medical
Engineering (CME), vol. 1, no. 1, pp. 788–792, 2012.
 
(4) I. Dan, T. Sano, H. Dan and E. Watanabe, “Optimizing the general linear model for
functional near-infrared spectroscopy: an adaptive hemodynamic response function
approach,” Neurophoton, vol. 1, no. 1, pp. 015004–015004, 2014.