Pelvic organ prolapse (POP) is one of the most common pelvic floor disorders (PFD) in females [1],[2]. Recent advances have been able to image the female pelvic floor using three dimensional (3D) transperineal ultrasound (TPUS) [3]. Previous works on 3D TPUS, have used deep learning methods to identify 2D locations including the plane of minimal hiatal dimensions (PMHD) and the levator hiatus. However, analysis of the entire 3D TPUS image still relies on manual segmentation. Thus, segmenting structures of interest using 3D deep learning methods can reduce the time required to analyze such images to provide a better understanding of the positions and interaction of the pelvic organs. The objective of this work was to develop a fully automated and first of its kind segmentation method to segment the pelvic organs in 3D TPUS to aid in the analysis and treatment of POP.

Segmentation models were developed using 3D TPUS images. The available dataset consisted of 161 3D TPUS volumes from 104 patients who presented to a tertiary urogynecology clinic with symptoms of pelvic floor disorders. Of the 161 3D TPUS volumes, 89 were captured at rest and 72 were captured during a Valsalva maneuver. An expert with prior experience in pelvic floor US interpretation and three trainees, who were trained by the expert, then manually segmented the relevant pelvic anatomical structures. These structures included the pubic symphysis, urethra, bladder, rectum, rectal ampulla, and anorectal angle. We separated the images patient-wise into train/validation/test sets with amounts 97/15/49 respectively. A deep learning segmentation model was developed using a 3D U-Net model with ResNet34 encoder. For comparison, a 2D slice-based U-Net model and a vanilla 3D U-Net model were also implemented. The models were compared to ground truth using the spatial overlap metric, i.e., the Dice similarity coefficient (DSC).

An example of one segmentation result from each tested model as well as the manual ground truth are shown in Figure 1. A full summary of results for each model can be seen in Table 1.

The proposed 3D U-Net with ResNet34 encoder had the greatest DSC for each of the segmented structures on the 49-image test set and could segment one TPUS image in less than five seconds. The proposed model achieved 53.9%, 32.1%, 67.1%, 53.9%, 15.9% and 34.3% DSC for pubic symphysis, urethra, bladder, rectum, rectal ampulla, and anorectal angle respectively, as shown in Table 1. The results demonstrated that a deep learning model can effectively segment pelvic organs in 3D TPUS.

Deep learning models can provide accurate segmentations at a fraction of the time required for manual segmentation. Allowing for easier and faster 3D analysis of the pelvic floor. While the DSC can be improved with better model selection and tuning, the current results can be used for making clinical measurements such as the bladder descent or rectal ampulla descent.

I. Nygaard et al., “Prevalence of symptomatic pelvic floor disorders in US women,” Jama, vol. 300, no. 11, pp. 1311–1316, 2008.
J. M. Wu et al., “Prevalence and trends of symptomatic pelvic floor disorders in us women,” Obstetrics Gynecol., vol. 123, no. 1, p. 141, 2014.
H. P. Dietz, “Pelvic floor ultrasound: A review,” Clin. Obstetrics Gynecol., vol. 60, no. 1, pp. 58–81, 2017.

Continence 7S1 (2023) 100997
DOI: 10.1016/j.cont.2023.100997