Automated Recognition and Measurement for Levator Hiatus in 3D Ultrasound

Objectives To evaluate the performance of a CNN‐based (convolutional neural networks‐based) AI software for automatic recognition and measurement of minimal levator hiatus in transperineal ultrasound volumes from postpartum women, and to assess its agreement with manual measurements in different functional states. Methods We conducted a retrospective analysis of 100 transperineal ultrasound volumes measured independently by two sonographers (one junior, one senior) using a SonoScape S60 ultrasound system. Manual measurements included anteroposterior diameter (AP), left‐to‐right diameter (LR), levator hiatus area (HA), which were assessed at rest, during maximum pelvic floor contraction maneuver, and during maximum Valsalva maneuver. The levator‐urethra gap (LUG) was additionally measured during contraction. The same volumes were identified and measured using automated CNN‐based software (Auto PF software, SonoScape). When automatic identification failed, manual slice adjustment was permitted before reattempting automated measurement. The automatic recognition rate was recorded in different functional states. Inter‐rater reliability between manual and automated measurements was evaluated using intraclass correlation coefficient (ICC) and Bland–Altman analysis. Results The overall automatic recognition rate was 86.69%, varying by functional state: 94% at rest, 86.17% during contraction, 79.80% during Valsalva. Automated measurements showed good agreement with manual measurements for HA, AP and LR (ICC > 0.75), excellent agreement with the senior sonographer's measurement for HA, AP (ICC > 0.90). The senior sonographer's measurements demonstrated higher concordance with automated results. Furthermore, the agreement was poorest for LUG measurements. Conclusions Automated software is feasible for the recognition and measurement of LH in postpartum women, which can reduce the inter‐observer variability, standardize pelvic floor assessment and improve workflow efficiency. Our study supports the clinical utility of CNN‐based AI ultrasound software, but optimization is needed for LUG measurements to enhance clinical applicability.