Artificial Intelligence to Determine Correct Midsagittal Plane in Dynamic Transperineal Ultrasound

Area of Science:

• Medical imaging
• Machine learning in healthcare
• Pelvic floor diagnostics

Background:

• Accurate capture of the midsagittal plane is crucial for dynamic transperineal ultrasound assessments of the pelvic floor.
• Current methods rely on examiner expertise, which can introduce variability.
• Objective tools are needed to standardize and improve the accuracy of midsagittal plane identification.

Purpose of the Study:

• To develop and validate a machine learning (ML) model for identifying the correct midsagittal plane in dynamic pelvic floor ultrasound.
• To assess the model's concordance with junior and senior examiners, using an expert examiner as the gold standard.

Main Methods:

• An observational, prospective study involving 90 patients without pelvic floor pathology.
• Ultrasound videos captured at rest and during the Valsalva maneuver were analyzed.
• A segmentation model, trained on prior data, was used with XGBoost algorithm and feature engineering to create the ML model. Concordance was measured using the kappa index.

Main Results:

• The ML model achieved a kappa index of 0.930 (p < 0.001), indicating very good agreement in detecting the correct midsagittal plane.
• A junior examiner showed very good agreement (kappa index = 0.930, p < 0.001).
• A senior examiner demonstrated good agreement (kappa index = 0.789, p < 0.001).

Conclusions:

• A validated machine learning model can accurately determine the midsagittal plane in dynamic transperineal ultrasound.
• The model's performance demonstrates a level of agreement comparable to or exceeding that of junior and senior examiners.
• This ML tool offers a promising approach for standardizing and enhancing the accuracy of pelvic floor ultrasound interpretation.