Segmentation of 3D Tubular Tree Structures in Medical Images ...

More documents

Recommendations

Info

70 Chapter 4. Tube Segmentation associated costs g(x) (the costs are computed as the average costs of the values computed at the discrete voxels). The utilized cost term g(x) incorporates the gradient magnitude |∇G σ ⋆ I| and a soft shape prior based on D surface that emphasizes edge information in proximity of the expected tube surface: g(x) = e − |∇Gσ⋆I(x)|2 2σ edge 2 ) (1 − αe − D surface (x)2 2σ 2 shape (4.1) where 0 ≤ α ≤ 1 can be used to control the influence of the shape prior. σ depends on the image noise level, while σ edge depends on the contrast and is application specific. The value of σ shape depends on the maximally expected variation from a perfectly tubular shape. 4.4 Experiments In this section we study the behavior of our previously presented tube segmentation methods with respect to their robustness against variations from a standard cylindrical tubular shape and disturbances in the image. Therefor we present qualitative and quantitative results achieved on clinical datasets showing such variations. Datasets and methods: The datasets we use for evaluation are an airway tree (Fig. 4.3 and 4.5) and four pathological abdominal aortas in contrast enhanced CT datasets (Fig. 4.4). The aortas had stenosis or aneurysms, as well as calcifications (Fig. 4.4(a)), thus their shape deviated significantly from a standard tubular shape. For each of these abdominal aorta datasets, two semi-automatically generated reference segmentations in image regions around the aneurysms/stenosis were available; one segmentation following the inner aorta wall including the calcifications and one segmentation that excludes the calcifications. The reference segmentations only contain the three branches around the main bifurcation of the abdominal aorta. The structural representations required for the segmentation methods were obtained using the GVF-based TDF with the offset medialness function (Section 2.3.2) and the GVF-based grouping and linkage method (Section 3.3). However, as the reference segmentations of the aortas only contain the three main branches of the abdominal aorta, the corresponding branches from the structural representations were selected manually and the others were discarded, such that the structures of the resulting segmentations are comparable. For comparison, we also show the shape priors reconstructed from these structural
4.4. Experiments 71 representations by using inverse distance transformations based on the centerlinepoints and radius estimates. Qualitative results: The airway tree segmentations were obtained for the whole tree structure (Figs. 4.3(d)-(f)) and also for a selected path starting from the trachea leading to one of the distal ends of the airway tree (Fig. 4.3(c) and the green part in Figs. 4.3(d)-(f)). As can be seen from Figs. 4.3(d)-(f), the segmentations associated with the selected path do not leak severely into the side branches of the airway tree. Note that in these areas no edge information is present. With the graph cut based method this is not very surprising as the method utilizes a hard shape prior that avoids leakage, however, the GVF-based method also shows a comparable behavior and no hard shape prior is utilized. This shows some robustness of the GVF-based approach against leakage. In Fig. 4.5 the segmentation results are shown in axial image slices along the selected path from Fig. 4.3(c). As can be seen, the shape of the airway tree shows larger deviations from a tubular shape and the cross section profiles deviate significantly from a perfectly circular or purely convex shape. While the shape prior still deviates significantly from the actual airway shape, the segmented surfaces obtained with the graph cut and GVF-based methods both show a good correspondence with the image data. In Fig. 4.4 the segmentation of one of the diseased abdominal aorta datasets is shown. The dataset shows an aneurysm and a severe stenosis due to calcification. The segmentation results for our presented methods are shown in Fig. 4.4(c) and (d), showing a qualitative difference between the two methods. The segmentation result of the graph cut based method seems smoother and includes the calcifications, while the GVF-based method on the other hand excludes the calcifications. With such calcifications two distinct edges can be observed in proximity of the shape prior and the graph cut based method chooses between the two possible edges in a globally optimal way based on the preference given by the soft shape prior. The GVF-based method on the other hand evolves a front from the centerline outwards until the first distinct edge is reached, thus showing a more local segmentation behavior. We also want to point out that the TDFs were capable of coping with these deviations from a perfectly cylindrical tube shape, although these conditions deviate from the assumptions of the TDFs. This shows that these assumptions are not too restrictive to cope with the deviations found in clinical practice.
Page 1:
Graz University of Technology Insti
Page 5 and 6:
Abstract The segmentation of tubula
Page 7 and 8:
Kurzfassung Die Segmentierung tubul
Page 9:
Acknowledgements I am deeply gratef
Page 13 and 14:
Contents 1 Introduction 1 1.0.1 Req
Page 15:
CONTENTS xv 8 Conclusion and Outloo
Page 18 and 19:
xviii LIST OF FIGURES 4.2 Shape pri
Page 21:
List of Tables 3.1 Average centerli
Page 24 and 25:
2 Chapter 1. Introduction (a) Porta
Page 26 and 27:
4 Chapter 1. Introduction (a) Volum
Page 28 and 29:
6 Chapter 1. Introduction (a) Inacc
Page 30 and 31:
8 Chapter 1. Introduction features
Page 32 and 33:
10 Chapter 1. Introduction 1.1.2 Di
Page 34 and 35:
12 Chapter 1. Introduction Input vo
Page 36 and 37:
14 Chapter 1. Introduction part: Me
Page 38 and 39:
16 Chapter 1. Introduction in the l
Page 40 and 41:
18 Chapter 2. Extraction of Tubular
Page 42 and 43: 20 Chapter 2. Extraction of Tubular
Page 70 and 71: 48 Chapter 3. Grouping and Linkage
Page 88 and 89: 66 Chapter 4. Tube Segmentation Onl
Page 90 and 91: 68 Chapter 4. Tube Segmentation obt
Page 94 and 95: 72 Chapter 4. Tube Segmentation (a)
Page 96 and 97: 74 Chapter 4. Tube Segmentation Tab
Page 99 and 100: Chapter 5 Liver Vascular Tree Segme
Page 101 and 102: 5.2. Method 79 (a) MIP of the datas
Page 103 and 104: 5.3. Evaluation and Results 81 5.3.
Page 105 and 106: 5.3. Evaluation and Results 83 cent
Page 107 and 108: 5.3. Evaluation and Results 85 (a)
Page 115 and 116: 5.4. Discussion 93 results. Selle
Page 117: 5.5. Conclusion 95 a prior to const
Page 120 and 121: 98 Chapter 6. Coronary Artery Tree
Page 131 and 132: Chapter 7 Airway Tree Segmentation
Page 133 and 134: 7.2. Methods 111 shown in Fig. 7.1.
Page 135 and 136: 7.2. Methods 113 Additionally, the
Page 137 and 138: 7.3. Evaluation and Results 115 Gro
Page 139 and 140: 7.3. Evaluation and Results 117 Tab
Page 141 and 142: 7.4. Discussion 119 Figure 7.5: “
Page 143 and 144:
7.5. Conclusion 121 Comparison to o
Page 145 and 146:
7.5. Conclusion 123 Figure 7.6: Exa
Page 147 and 148:
Chapter 8 Conclusion and Outlook Co
Page 149 and 150:
8.1. Conclusion 127 arteries - info
Page 151:
8.2. Directions for Future Work 129
Page 154 and 155:
132 Chapter A. List of Publications
Page 157 and 158:
BIBLIOGRAPHY 135 Bibliography [1] A
Page 159 and 160:
BIBLIOGRAPHY 137 [21] Choyke, P., Y
Page 161 and 162:
BIBLIOGRAPHY 139 [41] Florin, C., P
Page 163 and 164:
BIBLIOGRAPHY 141 [63] Kitslaar, P.
Page 165 and 166:
BIBLIOGRAPHY 143 [85] Lindeberg, T.
Page 167 and 168:
BIBLIOGRAPHY 145 [108] O’Donnell,
Page 169 and 170:
BIBLIOGRAPHY 147 [129] Schmugge, S.
Page 171 and 172:
BIBLIOGRAPHY 149 [150] van Bemmel,
Page 173:
BIBLIOGRAPHY 151 [171] Yi, J. and R
show all

Segmentation of 3D Tubular Tree Structures in Medical Images ...

You also want an ePaper? Increase the reach of your titles

Delete template?

Save as template?