Table of Contents:

Benchmark

Benchmarking, such as the 2017 AAPM Thoracic Auto-segmentation Challenge [149], can be helpful to understand the advantages and disadvantages of each method and compare architectures. Figure 7.8 shows a visual result of one patient using a DL-based method from the challenge [12].

From the report of this challenge [149], there were seven participants who completed the online challenge. Five out of seven participants used DL-based methods. In addition to those reported DL-based methods participating in this challenge, a review of recent studies using this benchmark data was performed. The numerical results of different DL-based methods are listed as follows: There was not a single method or architecture that outperforms others in every organ structure. This highlights the inherent challenge of multi-organ segmentations. From Table 7.8, it seems that U-net-GAN has the best performance in esophagus, lungs, and spinal cord, yet gives heart contours

Visual result of one patient’s data via DL-based method

FIGURE 7.8 Visual result of one patient’s data via DL-based method: (a) shows the CT image in axial view, (b) and (c) show the manual contour and segmented contour in axial view, respectively; (d) and (e) show the manual contour and segmented contour in 3D view.

trailing behind other methods. The results in this table also seem to indicate that 3D methods tend to perform better than the 2D counterpart. This is studied further in Chapter 8. It is worth noting that, except for the esophagus, most methods reach the level of accuracy equivalent to interobserver variability within clinical use.

Conclusion

This chapter covered the current state-of-the-art DL-based auto-segmentation architecture used for auto-delineation. Detailed discussion was given to cover all proposed DL-based methods found in the literature. Comparisons were made to highlight the novelty of each architectural approach and contrasted the pros and cons of them. In general, multi-organ segmentation is a challenging topic and currently there is not a clear architectural solution. The 2017 AAPM Thoracic Autosegmentation Challenge can be used as a benchmark to show the relative performance of recent DL-based methods. It further demonstrates the difficulty of this problem. Although there is no single architecture that outperformed others for all organs, the data do provide some insights towards the superiority of certain methods.

Judging from the statistics of the cited works, there is a clear trend of using FCNs to perform end-to-end semantic segmentation for multi-organ automatic segmentation. Recently, GAN-based methods have been used to enhance the reliability of segmented contours. R-CNNs and hybrid methods have started to gain popularity in medical image segmentation, although only a few methods surveyed were applied to multi-organ segmentation.

DL-based multi-organ segmentation techniques represent a significant innovation in daily practices of radiation therapy workflow, expediting the segmentation process, enhancing contour consistency, and promoting compliance to delineation guidelines [4, 12, 52, 86, 88, 91, 96]. Furthermore, DL-based multi-organ segmentation could facilitate online adaptive radiotherapy to improve clinical outcomes.

TABLE 7.8

Comparison of the Results from DL-Based Methods Using Datasets from the 2017 AAPM Thoracic Auto-segmentation Challenge

Metric

Method

Esophagus

Heart

Left Lung

Right Lung

Spinal Cord

DSC

DCNN

0.72 ±0.10

0.93 ± 0.02

0.97 ± 0.02

0.97 ± 0.02

0.88 ± 0.037

Team Elekta 3D U-net [150]

0.72 ±0.10

0.93 ± 0.02

0.97 ± 0.02

0.97 ± 0.02

0.89 ± 0.04

Multi-class CNN

0.71 ±0.12

0.91 ±0.02

0.98 ± 0.02

0.97 ± 0.02

0.87 ±0.110

Team Mirada 2D ResNet

0.61 ±0.11

0.92 ± 0.02

0.96 ± 0.03

0.95 ± 0.05

0.85 ± 0.035

Team Beaumont 3D and 2D U-net

0.55 ± 0.20

0.85 ± 0.04

0.95 ± 0.03

0.96 ± 0.02

0.83 ± 0.080

Team WUSTL U-net-GAN [12]

0.75 ± 0.08

0.87 ± 0.05

0.97 ± 0.01

0.97 ±0.01

0.90 ± 0.04

MSD

(mm)

DCNN

2.23 ± 2.82

2.05 ± 0.62

0.74 ±0.31

1.08 ±0.54

0.73 ±0.21

Team Elekta 3D U-net [150]

2.34 ± 2.38

2.30 ± 0.49

0.59 ± 0.29

0.93 ± 0.57

0.66 ± 0.25

Multi-class CNN

2.08 ± 1.94

2.98 ± 0.93

0.62 ± 0.35

0.91 ±0.52

0.76 ± 0.60

Team Mirada 2D ResNet

2.48 ± 1.15

2.61 ± 0.69

2.90 ± 6.94

2.70 ± 4.84

1.03 ±0.84

Team Beaumont 3D and 2D U-net

13.10 ± 10.39

4.55 ± 1.59

1.22 ±0.61

1.13 ±0.49

2.10 ±2.49

Team WUSTL U-Net-GAN [12]

1.05 ±0.66

1.49 ±0.85

0.61 ±0.73

0.65 ± 0.53

0.38 ± 0.27

HD95

(mm)

DCNN

7.3+10.31

5.8 ± 1.98

2.9 ± 1.32

4.7 ± 2.50

2.0 ± 0.37

Team Elekta 3D U-net [150]

8.71 + 10.59

6.57 ± 1.50

2.10 ±0.94

3.96 ±2.85

1.89 ±0.63

Multi-class CNN

7.8 ±8.17

9.0 ± 4.29

2.3 ± 1.30

3.7 ± 2.08

2.0 ± 1.15

Team Mirada 2D ResNet

8.0 ± 3.80

8.8 ±5.31

7.8 ± 19.13

14.5 ± 34.4

2.3 ±0.50

Team Beaumont 3D and 2D U-net

37.0 ± 26.88

13.8 ±5.49

4.4 ±3.41

4.1 ±2.11

8.10 ± 10.72

Team WUSTL U-net-GAN [12]

4.52 ±3.81

4.58 ± 3.67

2.07 ± 1.93

2.50 ± 3.34

1.19 ±0.46

*Note: Methods not followed by a reference were from the 2017 AAPM Thoracic Auto-segmentation Challenge report [149].

Acknowledgments

This chapter was supported in part by the National Cancer Institute of the National Institutes of Health under Award Number R01CA215718, and Dunwoody Golf Club Prostate Cancer Research Award, a philanthropic award provided by the Winship Cancer Institute of Emory University.

 
Source
< Prev   CONTENTS   Source   Next >