Automated Machine Learning Segmentation of the Pelvic Floor in 3D Transperineal Ultrasound

Objectives: Pelvic organ prolapse (POP), abnormal descent of one or more pelvic organs, is a common and prevalent pelvic floor disorder (PFD). Recent advances have enabled imaging of the pelvic floor using three dimensional (3D) transperineal ultrasound (TPUS). Previous research has utilized deep learning methods to characterize 2D structures and measurements, including the plane of minimal hiatal dimensions (PMHD) and the levator hiatus. However, analysis of the entire 3D TPUS volume is currently dependent on manual segmentation. Thus, using 3D deep learning methods to automate 3D segmentation of pelvic floor structures can reduce the time and manual effort required to analyze such images. This will aid in providing a better understanding of 3D position, orientation, and interactions of the pelvic organs. The objective of this work was to develop a novel, fully automated segmentation method to identify the pelvic organs on 3D TPUS.

Methods: Segmentation models were developed using 3D TPUS volumes from 104 patients who presented to a tertiary urogynecology clinic with symptoms of PFD. An expert with prior experience in pelvic floor US interpretation and three trainees, performed manual segmentation of relevant pelvic anatomical structures. These structures included the pubic symphysis, urethra, bladder, rectum, rectal ampulla, and anorectal angle. We developed a deep learning segmentation model using a 3D U-Net with ResNet34 encoder. For comparison, a 2D slice-based U-Net model and a vanilla 3D U-Net model were implemented. The models were compared to ground truth using Dice similarity coefficient (DSC) which measures spatial overlap between manual and automated segmentations. Higher overlap means better results where a 100% DSC means the automated and manual results are identical.

Results: We used a total of 161 3D TPUS volumes where 89 volumes were captured at rest and 72 volumes were captured during a Valsalva maneuver. Volumes were then separated into training, validation, and testing datasets based on amounts 97/15/49 respectively. Training and validation data are used to train the model while the test data is used for evaluation only. An example of one segmentation result from each tested model are shown in Figure 1. The proposed model achieved the highest DSC for all structures on the test set, as shown in Table 1. Our proposed model can segment one image in less than five seconds. This is the first work of its kind so no minimum threshold exists for determining clinical significance with regards to DSC.

Conclusions: Deep learning models can provide accurate segmentations without an expert tediously labelling the images. 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. Future work will explore minimum accuracy requirements for the deep learning models.