Status Plus

abstract

232 - MANUAL AND AUTOMATIC 3D SEGMENTATION OF THE PUBIC PART OF THE LEVATOR ANI MUSCLE IN ULTRASOUND IMAGES

232

MANUAL AND AUTOMATIC 3D SEGMENTATION OF THE PUBIC PART OF THE LEVATOR ANI MUSCLE IN ULTRASOUND IMAGES

VAN DEN NOORT1, A. T. GROB 2, M. VAN STRALEN 3, C. H. SLUMP 1, C. VAN DER VAART 2;
1MIRA Institute for Biomedical Technology and Technical Medicine, Univ. of Twente, Enschede, Netherlands, 2Department of Reproductive Medicine and Gyneacology, Univ. Med. Ctr. Utrecht, Utrecht, Netherlands, 3Image Division, Univ. Med. Ctr. Utrecht, Utrecht, Netherlands.

Introduction: Offline analysis of 3D ultrasound images of the levator ani muscle (LAM) is currently mainly restricted to 2D reconstructions. Since the LAM is complex in shape, important structural and functional information may be missed in 2D analysis.
Objective: To develop a manual and automatic 3D segmentation of the pubic part of the LAM in ultrasound images.
Methods: First, we randomly selected 20 ultrasound recordings from a dataset of 280 nulliparous women at 12 weeks of gestation[1]. In this dataset the LAM was manually outlined by two independent observers to analyse the similarity of the segmentations, inter and intraobserver. Secondly, 30 additional ultrasound recordings were selected from the dataset and segmented by one observer.
The 50 segmentations were resampled with a dedicated point distribution model (PDM), Fig. 1. With these resampled manual segmentations active appearance models (AAM)[2] were trained by the leave-one-out principle. These were then used for automatic segmentation of the LAM.

Fig. 1. PDM of a 3D segmentation of the LAM
The similarity between two segmentations was analysed by measuring the mean absolute distance (MAD) and the Dice value between the segmentations. The MAD is the mean of the distances measured from points on the surface of a segmentation to the closest surface point of the other segmentation, Fig. 2. The Dice is a measure of overlap between two segmentations: when they are of the same size and fully overlap, Dice is 1 and when there is no overlap Dice is 0.

Fig.2. Manual vs. AAM segmentation, demonstrating overlap and absolute distance measures.

Results:

Resulting MAD and Dice for the inter- and intraobserver and automatic vs manual segmentation.

 

MAD(mm)

  

Dice(mm)

  
 

Inter obs.

Intra obs.

AAM obs.

Inter  obs.

Intra obs.

AAM obs.

Median

1.10

0.76

1.24

0.60

0.68

0.60

1st Quart.

0.94

0.63

0.99

0.54

0.63

0.57

3rd Quart

1.28

0.98

1.37

0.63

0.71

0.63

Min.

0.65

0.58

0.75

0.41

0.56

0.07

Max.

1.65

1.33

10.24

0.68

0.74

0.73

Results are summarized in the tabel. The inter- and intraobserver MAD represents only a 1-3 voxels mismatch on average, the Dice values are moderate. The automatic method approaches the accuracy of the observers, expressed by the similar MADs and Dices for the AAM. The AAM showed five mismatches, which we suspect was caused by deviating appearance of the feces.
Conclusions: The MAD values are good, a 1-3 voxel mismatch is to be expected. When one delineates an object by hand there will always be a small difference of a few pixels (voxels in 3D), between delineations. The moderate Dice values can be explained by the long thin shape of the LAM, the Dice is more sensitive to small mismatches in this case than it would be in case of e.g. a sphere.
In conclusion we were able to reproducibly segment the pubic part of the LAM manually in 3D. The automatic segmentation is very promising, considering the median MAD and Dice approach inter observer variability. Improved automatic segmentation, decreasing the failure rate, is currently under investigation. We expect that mismatches can be circumvented by semi-automatic segmentation.
References: [1]. Ultrasound Obstet Gynecol. 2013; 42(5):590-5. [2]. International Ultrasonics Symposium. 2015; 7329123