Keywords: Pelvic floor, Levator Ani, ultrasound, deep learning, segmentation
TL;DR: Deep learning methods like U-Net can automatically segment LAM from 3D EVUS images, achieving a mean Dice score of 0.86 and indicating potential for AI-based diagnostic tools to improve management of pelvic floor disorders.
Abstract: The Levator Ani Muscle (LAM) deficit is a common side effect of vaginal childbirth and is linked to pelvic organ prolapse (POP) and other pelvic floor complications. Diagnosis and treatment of these complications require imaging and examining the pelvic floor, which is a time-consuming process subject to operator variability. We proposed using deep learning (DL) methods to automatically segment LAM from 3D endovaginal ultrasound images (EVUS) to improve diagnostic accuracy and efficiency. Over one thousand images extracted from 3D EVUS data consisting of healthy subjects and patients with pelvic floor disorders were utilized for the automated LAM segmentation. U-Net, FD-U-Net, Attention U-Net were implemented with Dice and Intersection over Union (IoU) used for model performance evaluation. The U-Net based models had 0.84-0.86 mean Dice score, which demonstrated efficacy compared to literature in LAM segmentation. Our study demonstrated the feasibility of DL-segmentation using U-Net and its variants for automated LAM segmentation from 3D EVUS images and has potential of being implemented in AI-based diagnostic tools for improved management of pelvic floor disorders, especially in low socioeconomic regions.