Automatic segmentation of the female pelvic floor muscles on MRI for pelvic floor function assessment

Background: Pelvic organ prolapse (POP) is a pelvic floor dysfunction disease which affects females. The volume of pelvic floor muscle, especially the levator ani muscle (LAM), is an important indicator of pelvic floor function. However, muscle volume measurements depend on manual segmentation, which is clinically time-consuming. In this work, we present an efficient automatic segmentation model of pelvic floor muscles with magnetic resonance imaging (MRI) based on DenseUnet, to achieve muscle volume calculation and provide a reference for the assessment of pelvic floor function. Methods: A total of 49 female pelvic floor magnetic resonance (MR) series were retrospectively enrolled from the First Affiliated Hospital of Army Military Medical University between 2013 and 2021, including 21 normal participants and 28 patients with stage 1-4 POP. The LAM, internal obturator muscle (IOM), and external anal sphincter (EAS) were manually segmented. An improved DenseUnet was proposed for automatic segmentation of these 3 muscles. The Dice similarity coefficient (DSC), Hausdorff distance (HD), and average symmetrical surface distance (ASSD) were used to evaluate segmentation results. The segmentation performance of the improved DenseUnet was compared with those of standard DenseUnet, ResUnet, Unet++, and Unet. Results: The improved DenseUnet showed a good performance. The average DSC and standard deviation of the LAM, IOM, and EAS was 0.758±0.151, 0.716±0.173, and 0.810±0.147, respectively. The average HD was 22.41, 19.00, and 36.01 mm, respectively; and the average ASSD was 3.66, 3.80, and 5.23 mm, respectively. The average DSC and standard deviation of the normal group and POP group was 0.779±0.166 and 0.757±0.154, respectively. There was no significant difference between the muscle volume of the improved DenseUnet and manual segmentation (all P values >0.05). The average total segmentation time for 1 case was 10.18 s on our setup, which is much lower than the manual segmentation time of 45 minutes. Conclusions: The improved DenseUnet segments the pelvic floor muscles in MRI quickly and efficiently, with good precision and faster speed than those of manual segmentation. This can assist doctors in quickly segmenting pelvic floor muscles, calculating muscle volume, and further evaluating pelvic floor function.