「Image Analysis of Diffusion Tensor MRI for White Matter Connectivity in the Human Brain」
（ＭＲテンソル画像の画像解析に関する研究）
We propose a method for detecting nerve fiber bundles in white matter
using diffusion tensor maps and apply the method to the practical
data. This dissertation is essentially dedicated to make a structural
brain maps about white matter fiber bundles and suggest the usefulness
for analyses of brain function.
Next, we consider a scalar potential field using several
characteristics of the diffusion tensor extracted from DTI. In such
field, we could give the high potential value for the high possibility
of nerve bundle existence. Then we propose the method to estimate the
connectivity of the nerve bundles between any two points in the field.
We also apply the method to the clinical data. From the result, we
obtain the line integral of the potential between two points on the
field that will be treated as the communication pathway of the nerve
bundles.
At last we show an novel method for mapping white matter connectivity
using diffusion tensor distance and also apply to in vivo brain
measurements. In the proposed method, we use directional diffusion
measurements to infer regional white matter connectivity. To assess
the connectivity, we compose the map based on the projected tensor
distance, then we put a label on the constructed map and segment
regionally connected white matter using labels.
（日本語訳）
本研究では，MR拡散テンソル画像から得られる複数の情報を組み合わせて，脳
内白質神経線維束を抽出するための新しい手法を提案し，この手法を実データ
に適用した結果について述べる．まず初めに，in vivoにおける脳の構造と機
能を定量的に解析するための一手法として，脳内白質神経線維束の強調と抽出
について述べる．
次に拡散テンソル画像に含まれるいくつかのテンソルの性質を利用したスカラ
ポテンシャルフィールドの合成について考える．このフィールド上では神経線
維束の存在が確からしいと考えられる部分に高いポテンシャルが割り当てられ
ている．その上で，フィールド上の任意の二点間の神経線維束の結合度を推定
する手法を提案する．またこの手法を実際の臨床データに適用した結果を示す．
最後に拡散テンソル距離を用いた白質神経線維束の新たな強調抽出手法とその適
用について述べる．提案手法では単位方向ごとに射影された，拡散テンソル距離
を用いる．神経線維束の結合度を評価するために，射影された拡散テンソルの距
離でラベリングされたマップをまず作成し，そのラベルに基づいたセグメンテー
ション処理を行なった．この手法により結合度にしたがった定量的な脳内白質神
経線維束を反映したマップを得ることができた．
「Microphone Array Steering Integration for Distant Talking Speech」
This thesis focuses on microphone array steering integration for distant
talking speech capture and recognition with a microphone array in noisy
reverberant environments.
Capturing and recognizing distant talking speech is indispensable for
achieving a teleconference system or speech control machine. However, the
quality of such speech can be degraded with a single transducer as a result
of room reverberation and ambient noise. A microphone array is an ideal
candidate to cope with this problem. However, its use requires target talker
localization and beamforming. To overcome these problems, we propose
localizing the distant target talker, capturing the talking speech with a
high quality, and recognizing the speech effectively with a microphone
array.
However, research on distant talking speech signal processing with a
microphone array has had problems in that they have required a lot of work,
because no common database with a microphone array had existed until now. We
have designed the Sound Scene Database in Real Acoustical Environments with
members (including the author) of the Real Acoustic Environments Working
Group in RWCP (Real World Computing Partnership). This database includes
some distant talking speech and
some impulse responses involving a microphone array in various environments.
Therefore, it is expected to be very useful for research with microphone
arrays from now on.
By using the above database, our focus in on multiple sound source
localization and talker localization. Until now, multiple sound source
localization has been difficult due to the effect of cross correlation in
multiple sound sources environments, although some sound source localization
algorithms have been proposed. To overcome this problem, we propose a
multiple sound source localization algorithm with the synchronous addition
of CSP (Cross-power Spectrum Phase analysis)
coefficients. The algorithm effectively localizes multiple sound sources.
However, it is difficult to localize the talker position among estimated
multiple sound source positions, even when the multiple sound sources are
localized effectively. Accordingly, we study talker localization among
estimated multiple sound source positions by the identification of speech or
non-speech with statistical speech or non-speech models, which are Gaussian
Mixture Models. This approach effectively localizes talker position among
estimated multiple sound source positions.
Next, capturing distant talking speech requires beamforming after talker
localization. Therefore, we focus on multiple beamforming as steering the
directivity of the microphone array. Multiple beamforming that steers the
directivity not only to the direct path but also to the reflection paths,
was proposed by J.L. Flanagan, et al. in 1993. However, to achieve effective
multiple beamforming in
real environments requires: 1). direct and reflection path localization, 2).
judgment on reflection signal utilization as target signals, and 3)
reflection signal equalization from distortion caused by wall impedances.
Accordingly, we investigate direct path and reflection path localization and
reflection signal utilization in various environments to solve 1) and 2).
Evaluation experiments confirm that direct and reflection paths can be
localized effectively, and the multiple beamforming performance ca be better
when the reverberation time in the room is large. We also propose a
reflection signal equalization algorithm with a cross-spectrum to solve 3).
Evaluation experiments confirm that multiple beamforming with equalization
is more effective than multiple beamforming without
equalization.
Finally, we try to design an automatic steering system incorporating a
microphone array and video camera based on the above research results for
capturing distant talking speech and talker images. The result is a
real-time automatic steering system.
「A Study on Three-Dimensional Visualization of Medical Image Information」
（医用画像情報の3次元可視化に関する研究）
The aim of this research is to develop an interactive
three-dimensional visualization method of medical image information
to improve intuitive understanding of condition of patients.
Moreover, three-dimensional visualized images, which have
high objectivity, promote smooth communication among doctors
or between doctors and patients, which is indispensable for
development of tele-medicine and informed consent.
The information from medical images divides broadly into two
categories: one is morphological and the other is functional.
In this paper, the author proposes the interactive visualization
methods to visualize both morphological and functional information.
For visualization of morphological information, enhancement of
morphological information and utilization of immersive projection
display enable users to grasp complex three-dimensional structure with
ease. On the other hand, blood flow pattern, which is one of the most
usable functional information, is visualized by interactive
application of multiple visualization methods.
The registration of heterogeneous measurements is executed and proved
to make it easy to recognize the fact which is difficult to grasp with
a single measurement. Brain walk-through environment, into which
structures of inner brain, skin and vessel ramification are integrated,
is developed and prove effectiveness of the registration.
本研究の目的は、医用画像情報のインタラクティブ性を備えた
3次元可視化手法を開発することである。
対象の本来の姿である3次元情報を医師の要求に即応した形で
提示することにより、2次元断層像と比較してより直観的に対象を
理解することが可能となる。
また客観性の高い3次元像の存在は複数の医師間および非専門家との
間の意志疎通を容易にするため、遠隔診断の発展やインフォームド
コンセントに大きく貢献することが期待される。
本論文では医用画像から得られる2種類の情報、すなわち形態情報と
機能情報の双方を対象として、インタラクティブ性を保持した
可視化手法を提案する。
形態情報の可視化に際しては、ユーザの目的に合致する情報を選別する
ことにより形態情報を鮮明化し、没入環境での提示を行うことで複雑な
立体構造の把握を可能とする。具体的な事例として血管が入り組んだ
複雑な構造を持つ腎臓糸球体の可視化を取り上げる。
一方、機能情報の可視化対象として、医療にとって非常に実用性が高く
かつ従来の2次元可視化では細部にわたる動態把握が困難であった血流を扱う。
カラードプラ法で取得した周期的に変動する血流の速度場を対象として、
異なる利点を持った複数の可視化手法を目的に応じてインタラクティブに
適用可能な3次元流れの可視化法を提案する。
更に、複数の可視化結果の統合提示を行うことにより、個別の可視化では
把握が困難であった事象の関連性を認識することが容易になる。
頭部の外部構造、内部から取得した構造および血流情報から得られる
血管走行の構造とをシームレスに統合した脳内ウォークスルー環境を構築し、
統合提示の有用性を示す。