Results

Atlas Ranking and Selection

One leave-one-out test for the head and neck cancer patients was used to illustrate the efficiency of atlas ranking and selection. The test patient had a large air pocket inside the esophagus, which made it appear different from most atlases. The first phase of atlas selection identified 12 atlases, and their contours were deformed to the test patient. These deformed contours demonstrate various agreement with the manual contour, as shown in Figure 5.5. In the second phase of atlas selection, the symmetric KL divergence of local-intensity histograms ranked atlases 12, 8, and 10 as the first, second, and third best atlases, which was consistent with the observation, as shown in Figure 5.5. The overlap ratio analysis for the 12 atlases is shown in Figure 5.4. In this analysis, a subset of five optimal atlases was further identified for contour fusion on the basis of the atlas selection rules. By comparing the auto-segmented contour generated from the five optimal atlases with the manual contour, a DSC of 0.67 and a MSD of 2.3 mm was found; the segmentation generated from the 12 atlases, without performing a second phase of atlas selection, had values of 0.31 mm and 4.0 mm, respectively. A subjective comparison is illustrated in Figure 5.6. It shows much better agreement between the auto-segmented contour and the manual contour for the approach using atlas selection than for the approach without atlas selection.

Esophagus Segmentation for Head and Neck Cancer Patients

Table 5.2 shows the results of the leave-one-out tests of the proposed MAS-AS approach compared with the MAS approach, as applied to 21 head and neck cancer patients. The MAS-AS approach resulted in a mean DSC of 0.70 ± 0.07 and a mean MSD of 1.9 mm ± 0.7 mm. For the MAS approach, patients 14-21 were used as a fixed set of atlases to delineate the esophagus for the first 13 patients. In these 13 tests, the MAS approach achieved a mean DSC of 0.55 ± 0.15 and a mean MSD of 2.5 mm ± 0.9 mm. The results for the first 13 patients were tested to determine whether a statistically significant difference existed between the MAS-AS and MAS approaches. Significantly different DSCs and MSDs were found for these two approaches (p-value of 0.005 for DSC and 0.022 for MSD). The mean DSC and mean MSD for the MAS-AS approach applied to the first 13 patients were 0.69 ± 0.08 and 2.0 mm ± 0.7 mm, respectively. These results were better than those of MAS approach. The DSCs and MSDs for patients 2,6,9, 10, and 11 showed significant improvement with the MAS-AS approach. However, it was observed that the MAS-AS approach was slightly inferior to the MAS approach in some patients in terms of the DSC and MSD evaluation. In addition, the

Esophagus contours of 12 atlases were deformed to the test image

FIGURE 5.5 Esophagus contours of 12 atlases were deformed to the test image. The red and green colors indicate deformed and manual contours, respectively. In the second phase of atlas selection, atlases 12, 8, and 10 were chosen as the first, second, and third best atlases, which is consistent with the observation; they have a better agreement between the manual contour and the deform contour than other atlases do.

Comparison of auto-segmentation using five selected optimal atlases

FIGURE 5.6 Comparison of auto-segmentation using five selected optimal atlases (red contour) versus that using 12 atlases without the second phase of atlas selection (yellow contour). Green colorwash indicates the manual contour, (a) Axial view, (b) Sagittal view.

TABLE 5.2

Segmentation Results of the Proposed Multi-Atlas Segmentation with the Online Atlas Selection (MAS-AS) Approach Compared with the Multi-Atlas Segmentation with a Fixed Set of Atlases (MAS)

Patient No.

Dice Similarity Coefficient

Mean Surface Distance (mm)

MAS-AS

MAS

Single Atlas

MAS-AS

MAS

Single Atlas

1

0.61

0.56

0.61

2.1

2.7

2.1

2

0.67

0.41

0.70

2.4

3.2

2.1

3

0.65

0.56

0.61

2.1

2.2

2.0

4

0.77

0.78

0.69

1.6

1.5

3.0

5

0.56

0.40

0.63

3.7

3.9

3.1

6

0.78

0.49

0.71

1.1

1.5

1.5

7

0.74

0.74

0.71

2.1

2.1

2.2

8

0.61

0.63

0.60

2.9

2.6

2.0

9

0.76

0.54

0.68

1.4

2.2

2.1

10

0.67

0.22

0.72

2.3

4.5

1.9

11

0.72

0.60

0.74

1.5

2.1

1.4

12

0.59

0.53

0.64

1.9

2.3

1.5

13

0.77

0.73

0.91

1.3

1.4

0.6

14

0.81

-

0.74

1.3

-

1.8

15

0.72

-

0.73

1.6

-

1.7

16

0.65

-

0.56

2.6

-

3.4

17

0.70

-

0.67

1.7

-

1.7

18

0.79

-

0.78

1.6

-

1.3

19

0.69

-

0.44

1.8

-

3.5

20

0.78

-

0.70

1.1

-

1.7

21

0.65

-

0.57

2.7

-

3.3

Median

0.70

0.56

0.69

1.8

2.2

2.0

Mean

0.70

0.55

0.67

1.9

2.5

2.1

SD

0.07

0.15

0.09

0.7

0.9

0.8

Note: Approach in 21 head and neck cancer patients. The column of “Single Atlas” reported the best result from a single atlas-based segmentation, which serves for the validation of deformable registration algorithm used in MAS-AS. The MAS approach used patients 14-21 as atlases to delineate the esophagus for the first 13 patients. SD = Standard deviation.

best overlap value from a single atlas in order to validate the deformable registration algorithm used in the proposed MAS-AS approach is reported in Table 5.2. The overlap values were measured between the deformed contour from each atlas and the best results were reported. These results showed reasonable agreements between the manual contour and the deformed contour thus justifying the use of the deformable registration algorithm.

Validation with Public Benchmark Dataset

Table 5.3 shows the results of the MAS-AS approach validated on the 24 test cases from the 2017 AAPM Thoracic Auto-segmentation Challenge dataset, which includes contour data of the esophagus, spinal cord, left lung, right lung, and heart. For the esophagus, the DSC ranged from 0.31 to 0.83 with a median of 0.69 and an average of 0.66 ±0.12. The MSD values ranged from 1.24 mm to 6.11 mm with a median of 2.21 mm and an average of 2.59 mm ± 1.28 mm.

TABLE 5.3

Segmentation Results of the Proposed Multi-Atlas Segmentation Method Evaluated on 24 Thoracic Cancer Patients Used in the 2017 AAPM Thoracic Auto-segmentation Challenge

Patient No.

Esophagus

Spinal Cord

DSC

MSD

HD95

DSC

MSD

HD95

1

0.50

3.88

15.54

0.86

0.79

2.42

2

0.60

2.65

12.68

0.91

0.58

2.02

3

0.71

1.96

6.57

0.83

0.95

2.95

4

0.31

5.94

21.15

0.88

0.85

2.37

5

0.74

1.65

6.23

0.80

0.80

2.07

6

0.76

1.28

4.34

0.84

0.89

2.67

7

0.68

2.46

11.42

0.81

1.23

3.31

8

0.79

1.35

4.35

0.80

1.00

2.98

9

0.63

2.31

6.45

0.63

0.64

1.76

10

0.74

1.79

5.41

0.82

0.88

2.65

11

0.65

2.19

9.41

0.75

0.95

2.47

12

0.60

2.98

10.42

0.80

0.74

2.42

13

0.63

3.10

21.12

0.89

0.54

1.52

14

0.76

1.49

4.44

0.85

0.81

3.19

15

0.70

2.24

10.22

0.78

0.65

2.00

16

0.74

1.77

7.07

0.87

0.85

.3.18

17

0.50

6.1 1

28.68

0.85

0.57

2.00

18

0.71

2.10

11.85

0.86

0.72

2.26

19

0.83

1.24

3.90

0.87

0.72

1.95

20

0.66

2.89

15.66

0.85

0.90

2.94

21

0.75

1.79

5.49

0.69

0.78

2.01

22

0.74

2.15

7.58

0.79

0.74

1.98

23

0.55

3.35

14.11

0.70

0.96

2.45

24

0.60

3.44

15.09

0.71

0.82

2.33

Median

0.69

2.21

9.81

0.83

0.80

2.40

Mean

0.66

2.59

10.80

0.81

0.81

2.41

SD

0.12

1.28

6.35

0.07

0.16

0.49

(Continued)

TABLE 5.3 (CONTINUED)

Segmentation Results of the Proposed Multi-Atlas Segmentation Method Evaluated on 24 Thoracic Cancer Patients Used in the 2017 AAPM Thoracic Auto-segmentation Challenge

Patient No.

Left Lung

Right Lung

Heart

DSC

MSD

HD95

DSC

MSD

HD95

DSC

MSD

HD95

1

0.97

0.95

3.17

0.96

1.47

13.55

0.94

2.60

6.84

2

0.95

1.59

11.38

0.96

1.74

9.90

0.90

3.98

15.70

3

0.97

1.41

9.79

0.97

2.15

18.28

0.91

3.48

12.43

4

0.98

1.20

6.16

0.98

0.80

2.75

0.91

3.71

14.30

5

0.90

3.45

28.40

0.97

1.24

3.96

0.91

3.55

10.87

6

0.93

3.12

19.08

0.96

2.17

9.73

0.94

1.77

5.15

7

0.94

2.60

11.16

0.87

4.30

34.98

0.90

3.41

10.66

8

0.96

1.64

7.08

0.96

1.75

6.78

0.90

3.83

11.48

9

0.95

2.41

7.96

0.92

3.66

24.74

0.91

4.00

9.68

10

0.98

III

5.83

0.98

0.78

2.98

0.91

3.54

12.35

II

0.81

6.15

51.27

0.95

2.42

10.52

0.91

4.27

11.07

12

0.97

2.00

9.91

0.97

1.21

6.81

0.90

4.73

14.19

13

0.98

1.00

6.88

0.98

1.30

9.07

0.92

2.70

10.04

14

0.97

1.12

4.42

0.98

1.09

5.71

0.91

3.68

10.09

15

0.97

1.34

10.39

0.97

1.71

15.36

0.92

3.58

11.98

16

0.97

1.63

12.89

0.97

1.91

18.15

0.91

3.01

9.59

17

0.95

2.85

10.95

0.95

2.20

12.54

0.88

4.26

16.79

18

0.95

2.08

10.85

0.93

4.91

28.32

0.88

5.35

31.39

19

0.96

1.53

6.07

0.91

2.63

18.13

0.90

3.19

8.98

20

0.95

2.45

12.31

0.87

3.81

23.77

0.89

3.62

15.52

21

0.79

6.38

35.92

0.97

1.83

6.34

0.82

7.77

31.77

22

0.95

2.53

18.62

0.96

2.84

18.30

0.87

5.11

18.82

23

0.76

9.12

86.43

0.97

1.47

7.17

0.89

5.19

24.79

24

0.98

1.16

6.31

0.97

1.83

17.42

0.95

2.09

6.69

Median

0.96

1.82

10.62

0.96

1.83

11.53

0.91

3.65

11.73

Mean

0.94

2.53

16.38

0.95

2.13

13.55

0.90

3.85

13.80

SD

0.06

2.00

18.55

0.03

1.08

8.44

0.03

1.22

6.89

Note: Results are shown for the esophagus, spinal cord, left lung, right lung, and heart. DSC = Dice similarity coefficient. MSD = Mean surface distance (mm), HD95 = 95% Hausdorff distance. SD = Standard deviation.

Illustration of auto-segmented contours

FIGURE 5.7 Illustration of auto-segmented contours (lines) compared with manual contours (colorwash) for one thoracic patient from the public benchmark dataset in several slices. The following organs are included: left lung (dark green), right lung (cyan), heart (magenta), esophagus (green), and spinal cord (red).

The 95% HD values ranged from 3.90 mm to 28.68 mm with a median of 9.81 mm and an average of 10.80 mm ± 6.35 mm.

For the spinal cord, the DSC ranged from 0.63 to 0.91 with a median of 0.83 and an average of

0.81 ± 0.07. The MSD values ranged from 0.54 mm to 1.23 mm with a median of 0.80 mm and an average of 0.81 mm ± 0.16 mm. The 95% HD values ranged from 1.52 mm to 3.31 mm with a median of 2.40 mm and an average of 2.41 mm ± 0.49 mm.

For the right lung, the DSC ranged from 0.87 to 0.98 with a median of 0.96 and an average of 0.95 ± 0.03. The MSD values ranged from 0.78 mm to 4.91 mm with a median of 1.83 mm and an average of 2.13 mm ± 1.08 mm. The 95% HD values ranged from 2.75 mm to 34.98 mm with a median of 11.53 mm and an average of 13.55 mm ± 8.44 mm. For the left lung, the DSC ranged from 0.76 to 0.98 with a median of 0.96 and an average of 0.94 ± 0.06. The MSD values ranged from 0.95 mm to 9.12 mm with a median of 1.82 mm and an average of 2.53 mm ± 2.00 mm. The 95% HD values ranged from 3.17 mm to 86.43 mm with a median of 10.62 mm and an average of 16.38 mm ± 18.55 mm. The large HD value was mainly due to the tumor exclusion. Tumors were excluded from the manual lung contours, while atlas-based segmentation, in general, is not able to handle this exclusion correctly due to different tumor locations on different patients.

For the heart, the DSC ranged from 0.82 to 0.95 with a median of 0.91 and an average of 0.90 ± 0.03. The MSD values ranged from 1.77 mm to 7.77 mm with a median of 3.65 mm and an average of 3.85 mm ± 1.22 mm. The 95% HD values ranged from 5.15 mm to 31.77 mm with a median of

11.73 mm and an average of 13.80 mm ± 6.89 mm.

Figure 5.7 illustrates the comparison of the auto-segmented contours with the manual contour for one patient in several slices. Good agreement was observed for all the structures under consideration, though with the esophagus having more variations than other structures.

 
Source
< Prev   CONTENTS   Source   Next >