서지주요정보
Cortical Surface Segmentation and Topology Correction for 3D Visualization of Brain MR Images = 뇌자기공명영상의 3차원 가시화를 위한 피질표면의 분할과 위상보정
서명 / 저자 Cortical Surface Segmentation and Topology Correction for 3D Visualization of Brain MR Images = 뇌자기공명영상의 3차원 가시화를 위한 피질표면의 분할과 위상보정 / Jin-Young Hwang.
발행사항 [대전 : 한국과학기술원, 2011].
Online Access 원문보기 원문인쇄

소장정보

등록번호

8022966

소장위치/청구기호

학술문화관(문화관)B1층 보존서고

DEE 11045

휴대폰 전송

도서상태

이용가능(대출불가)

사유안내

반납예정일

리뷰정보

초록정보

Human brain magnetic resonance (MR) images are widely used for the several purposes, such as visualization of the cortical surface, measurement of the cortical thickness, generation and degeneration of the brain, and diagnosis and analysis of disease. In these types of studies, a brain region is extracted from MR images in an important preprocessing step called skull stripping. In general, skull stripping is performed either by manual segmentation or automatic segmentation. However, because manual segmentation has drawbacks, such as the need for trained experts and the laboriousness of the task, automatic skull tripping is generally preferable in large-scale studies. After the skull stripping, we can visualize the brain region, including the cerebrum, the cerebellum, and part of the brain stem. Only the cerebrum volume is needed to visualize the cerebral cortex, so the cerebellum and the brain stem should be removed. By using the cerebellum segment, we can investigate various functional activities of the cerebellum, such as the motor, balance, language, and distance measurement functions. A skull-stripped volume without the cerebellum is then obtained, and the cerebral cortex can be transformed into a sphere or flat map for analysis of the cerebral cortex across populations. However, the cerebrum volume has topological defects, such as tunnels and holes, which prevent the occurrence of homeomorphism when the cerebral cortex is transformed into a sphere or flat map. This thesis proposes a skull-stripping method based on a level set method with a speedup operator and a refinement process. We first extend a conventional 2D level set method to a 3D equation and then apply the speedup operator to reduce the computation load. The speedup operator can improve the propagation speed of zero level surfaces while maintaining a signed distance function. Because the skull-stripped volume includes the cerebrum, the cerebellum, and part of the brain stem, we propose a cerebellum segmentation method based on an active contour model with a prior shape. We then obtain the cerebrum volume by extracting the cerebellum. Finally, to correct the topological defects of the cerebrum volume, we propose a topological correction method based on a graph theory. The experimental results were obtained with simulation data and in vivo brain images acquired from 1.5T and 3T MRI systems. The speedup operator reduces the total iterations of the synthesized volume by 75%. The proposed method was applied to several datasets and compared with existing methods such as BrainVisa, BET, and FreeSurfer. The proposed method provides a Jaccard index results of 0.971±0.0114 for the BrainWeb dataset, 0.864±0.035 for the IBSR dataset, and 0.9414±0.0517 for a self-produced dataset acquired from the 3T MRI system. In addition, the segmented cerebellum is compared with the FreeSurfer results. Using a manually segmented cerebellum as a reference, we obtained the following average Jaccard coefficients for the proposed method: 0.882 for the BrainWeb dataset and 0.885 for the 3T MRI dataset. Finally, the proposed topological correction method efficiently and intuitively provides an image of the cerebral cortex without any defects.

뇌자기공명영상(뇌 영상)은 뇌 기능과 대뇌의 다중 가시화, 대뇌 피질 표면의 두께 측정, 뇌의 성장과 퇴화, 뇌 관련 질병의 진단 및 분석에 많이 사용되고 있고, 이러한 연구는 뇌 영상으로부터 뇌 영역을 분할하는 단계에서 시작된다. 뇌 영상에 대하여 뇌 영역 추출(brain region extraction)을 적용하면, 대뇌뿐 아니라 소뇌와 뇌줄기(brainstem)가 포함된 뇌 영역을 얻을 수 있는데, 대뇌 피질 표면을 분석하기 위해서는 소뇌 분할 알고리즘으로 소뇌를 제거해 주어야 한다. 다음 단계로, 여러 뇌 영상들을 분석하기 위해 대뇌를 구 혹은 2차원 평면으로 변환해야 하는데, 이때, 대뇌 피질 표면에 위상 오류가 있으면, homeomorphism을 만족하지 못하여 왜곡된 결과를 얻을 수 있다. 따라서, 대뇌 피질 표면의 위상 오류를 보정하여, 대뇌가 구와 같은 위상을 갖도록 해야 한다. 본 논문에서는, 첫째로, 속도 연산자를 가지는 3차원 레벨셋을 이용한 뇌 영역 추출을 제안하였다. 슬라이스 간 영향을 고려하기 위하여 기존의 2차원 레벨셋을 3차원 레벨셋으로 확장하고, 반복 연산을 줄이기 위하여 속도 연산자를 적용하여 제로 레벨 커브가 빠르게 전파되도록 하였다. 다양한 뇌 영상을 이용하여 기존에 제안된 다른 방법들과 비교실험을 하였다. 실험 결과, 반복 연산 횟수를 75% 가량 줄여 3D 레벨셋의 문제인 계산 복잡도를 개선하였고, 뇌 영역 분할 결과 역시 다른 방법들과 비교하였을 때 다양한 뇌 영상에 대하여 더욱 안정적인 결과를 얻을 수 있었다. 둘째로, 뇌 영역 추출을 통하여 얻은 뇌 영역으로부터, 형태 제한을 하는 동적 윤곽선 모델(active contour model)을 이용하여 소뇌를 분할하는 방법을 제안하였다. 이 방법은 형태 제한이, 복잡하거나 모호한 경계를 효과적으로 구분할 수 있음을 확인하였고, 여러 뇌 영상에 적용하였을 때, 안정된 결과를 얻을 수 있었다. 셋째로, 대뇌 피질 표면의 위상 오류를 보정하기 위한 방법을 제안하였다. 이 방법은 일반적인 그래프 이론에 기반을 둔 것으로, 표면에 존재하는 터널(tunnel)과 구멍(hole)의 크기를 고려하여 정해진 조건에 따라 채우거나 개방토록 하였다. 이렇게 함으로써, 대뇌 피질 표면의 변화를 최소화할 수 있었으며, 위상 오류가 없는 볼륨을 구할 수 있었다.

서지기타정보

서지기타정보
청구기호 {DEE 11045
형태사항 viii, 98 p. : 삽화 ; 30 cm
언어 영어
일반주기 저자명의 한글표기 : 황진영
지도교수의 영문표기 : Hyun-Wook Park
지도교수의 한글표기 : 박현욱
학위논문 학위논문(박사) - 한국과학기술원 : 전기및전자공학과,
서지주기 References : p.93-98
QR CODE

책소개

전체보기

목차

전체보기

이 주제의 인기대출도서

Inferior view of the brain, which consists of the cerebrum, cerebellum, and brainstem [1]). The cerebrum contains sulci and gyri. As depicted in this figure, the brainstem can be hardly difficult to segment from cerebrum and cerebellum.

Medial view of the brain [1]. The cerebellum is directly connected to the cerebrum and brainstem. Neighbored area near the occipital lobe and the cerebellum has ambiguous boundary due to the gray matter of each. Segmenting the cerebellum is a challenging issue.

Illustration of Broadmann area [6]. (Left) Inferior view of the brain. (Right) Medial view of the brain. Each number represents the specific area corresponding to brain functions. For example, BA4 is the primary motor cortex, BA17 is the primary visual cortex, and BA41 and BA42 correspond closely to primary auditory cortex.

Various imaging techniques for functional brain mapping [13]. Among them, fMRI is noninvasive method, and has potential to link high spatial and temporal resolution studies to an understanding of systems organization across the brain [13].

Illustration of fMRI hemodynamics [15]. As increment of neural activity, glucose and blood flow are also increased. Indeed, paramagnetic and diamagnetic spins are accompanied. These will cause the fMRI signal intensity.

Implicit representation ofthe curve. The outer region is definedbypositive, whereas the inner region is defined by negative. The boundary of the curve is 0.

Block diagram of the proposed segmentation algorithm, which contains skull stripping phase and cerebellum segmentation phase. Skull stripping method is described in section 2.2.2, and cerebellum segmentation is presented in section 2.2.3. In training phase is also described in section 2.2.3.

Flow chart of the proposed speedup operator.

Schematic diagram for speed-up function in the 3D level set. Every blocks on the interface of 2 are checked whether the advanced block (B*)along normal vector (Vk) is similar to the current block.

(a) Coronal image of 1.5T MRI data, and (b) Amagnified view of dotted box of(a). As shown in (b), the boundary between the internal carotid artery and the white matter within brain region is ambiguous due to the similar intensity.

Flow chart of the proposed cerebellum segmentation method.

Segmentation results of 3T MRI data from FreeSurfer. Some regions show wrong segmentation results, indicated by the white arrow.

The evaluation results of the speedup operator. (a) Three orthogonal images of 3T MRI data (#4). (b) The level set method with the speedup operator approaches the target much faster than that without the speedup operator.

(a) Segmentation result without the refinement process. The anterior region was not properly segmented. (b) Segmentation result with the refinement process. The under-segmented regions are properly extracted.

Results of the skull stripping. (a) BrainWeb - NOISEORFO (b) BrainWeb - NOISE3RF20 (c) BrainWeh NOISES RF20(d) IRSR (e) 5T(f) 3T

Results of skull-stripping for the BrainWeb data (9% noise and 40% RF non-uniformity), Left to right columns: original images, skull-stripped imagesfrom BrainVisa, BET. FreeSurfer, and the proposed method.

Results of skull-stripping for the IBSR data (205_3). Left to right columns: original images, skull- stripped images from BrainVisa, BET, FreeSurfer, and the proposed method.

Results of skull-stripping for the 3T MRI data. Left to right columns: original images, skull-stripped images from BrainVisa, BET, FreeSurfer, and the proposed method.

Performance comparison of BrainVisa, BET, FreeSurfer, and the proposed method for the BrainWeb dataset.

Performance comparison of BrainVisa, BET, FreeSurfer, and the proposed method for the IBSR dataset.

Performance comparison of BrainVisa, BET, FreeSurfer, and the proposed method for the 3T MRI data.

Intermediate results from (a) the initial contour to (b) sagittal, (c) transverse, and (d) sagittal-view images.

Segmentation results from the proposed cerebellum extraction algorithm. (a) BrainWeb data (0% noise and 0% RFnon-uniformity), (b) 3T MRI data (#1),(c) 3T MRI data (#2).

Segmentation results from the conventional active contour model. Since the boundary between the cerebellum and cerebrum is ambiguous, the algorithm can yield oversegmented results.

Evaluation results of various indices for the 3T MRI data (#3) with respect to the slice number.

Performance comparison ofFreeSurfer and the proposed method using BrainWeb data

Performance comparison ofFreeSurfer and the proposed method using five 3T MRI data.

Segmentation results of FreeSurfer (left images) and the proposed method (rightimages) using one of the 3T MRI datasets. The peduncle region indicated by the white arrow in (b) is over-segmented by FreeSurfer, whilst the proposed method provides the clear boundary ofthe cerebellum.

The illustration of topological features.

D6 and D18 rules. D6is used for the foreground, whereas D18 is used for the background

Foreground labeling. 0 means background, 1 means foreground object, and 2 means islands. In group of 1, 0's blocks represent holes. Simply, by taking 1's group, the foreground object, whose size is biggest, is obtained without islands.

The illustrations ofhole detection. (a) Island-removed volume. (b) Afterinverting volume (a), hole detection is applied to remove the hole inside the island-removed volume.

The illustration of definition of r, C, and d. C represents centroid of the hole's volume (hole), r represents the averaged distances from C to boundary of the hole's volume, and d is the distance between minimal distance from boundary ofthe hole's volume to cortical surface (input volume).

The illustrations of the various cases. (Top) d is very small comparing with the hole's size. Thus, cutting d is better choice than filling the hole. (Bottom) The opposite case of the top row. If d is large comparing with hole's size, fillingthe hole is better choice than cutting d.

Synthetic volume having 1 finger and 4 tunnels. Each component cannot be easily differentiated with others.

(a)2D tomography ofthe volume. The handle looks like an island. Iftheisland propagates upper and lower slices, itwill be finally connected. Thus the island represents a part ofhandle. (b) Graph representation of (a).

Teapot example. (a) Lateral view (b) Black line estimates the maximum tunnel in Z (superior to inferior) direction.

dhandle and dtunnel are calculated to decide whether the handle is tore or the tunnel is filled.

A synthesized cup with 1 finger, 1island, and 2 tunnels.

Experimental results of the proposed method using the volume in Figure 36. (a) Island is removed. whereas other components remain. (b) Two handle is torn, and finger remains. g of the resultant volume is 0.

A synthesized volume with 2 islands, 5 holes, and 1 tunnel. (a) Surface rendering without transparency. (b)Surface rendering with transparency

Experimental results ofFigure 38. (a) Two islands are removed. (b) and (c) Results of hole handling with transparency. Five holes are pulled out since we do not apply the decision rule. (d)Long handle is tore, and genus of the volume is 0. More detailed resultis provided in Figure 40.

17 slices ofFigure 39(d). There are some holes, which are connected to upper and lower slices. They are finally linked to the background.

Surface rendering of the experimental result. (a) The original skull-stripped brain volume. (b) Topology corrected brain volume.

Magnified view ofFigure 41. The tunnel is completely filled.

(a) White matter surface of the right hemisphere. (b) Topology corrected volume of (a).

Magnified view of Figure 43. (a) and (c) show the surface of the original volume. (b) and (d) show the surface oftopology corrected volume. Island in (a) are clearly removed, and tunnel in (b)is filled.