Quantitative Cystocele Assessment in Clinical Pelvic Floor Ultrasound Diagnosis

Cystocele is a pelvic floor dysfunction disease to which women are prone after childbirth. The accuracy of pelvic floor ultrasound as the most commonly used examination method is influenced by subjective factors such as doctor experience and fatigue level, making it challenging to achieve high accuracy, consistency, and repeatability of diagnosis. This study aims to propose a high-precision and fully automatic cystocele evaluation method based on pelvic floor ultrasound video images.

This study retrospectively collected pelvic floor ultrasound images of 158 female G1P1 (first gestation and first parturition) patients from 2020 to 2024. According to the ultrasound diagnosis made by two senior doctors as the standard, 81 cystoceles and 66 non-cystocele patients were enrolled. Firstly, the ResNet34-UNet was used for automatic urethra segmentation. Then, key points were generated based on the automatically extracted urethra centerline. Features such as urethral key point displacement, urethral curvature change, and urethral inclination angles and their change were extracted for patients between rest and maximum Valsalva states. The support vector machine (SVM) classification model was used for cystocele prediction.

This study constructed two classification models to predict cystocele. One extracted the above features based on the automatic urethra segmentation, while the other extracted them based on the doctor-annotated urethra. The experimental results show that both models have achieved good prediction results, with AUCs of 91.37% and 98.58%, respectively. Model performance based on the urethral image delineated by the doctor is better, with an AUC improvement of 7.21% based on the independent test set.

The proposed method can achieve high-precision, repeatable, fully automatic quantitative cystocele evaluation in pelvic floor ultrasound examinations.