[Deep learning models based on fused ultrasound images for the assessment of cystocele in women].

Authors

Abstract

Cystocele is a common type of pelvic organ prolapse in women. Ultrasound is the preferred imaging modality for the diagnosis of cystocele; however, the examination procedure is relatively complicated and dependent on the operator's experience, resulting in subjective interobserver differences in diagnosis. This study aims to develop deep learning models based on two-dimensional (2D) ultrasound images, three-dimensional (3D) ultrasound images, and fused 2D and 3D ultrasound images for the diagnosis and ultrasonographic grading of cystocele in women, and to compare the diagnostic performance of the 3 models. Data were retrospectively collected from women who underwent transperineal pelvic floor ultrasonography at Xiangya Hospital, Central South University, between January 2022 and December 2024. A total of 625 patients met the inclusion and exclusion criteria. According to the clinical and ultrasonographic diagnoses, 467 patients with cystocele, including 167 with mild cystocele and 300 with significant cystocele, and 158 patients without cystocele were ultimately included. A GE Voluson E10 ultrasound system was used to acquire paired 2D midsagittal images of the bladder at rest and during maximal Valsalva maneuver, as well as reconstructed images of the levator hiatus in 3D volume-rendering mode during maximal Valsalva maneuver. All data were randomly divided at a ratio of 7꞉3 into a training set comprising 437 patients and a test set comprising 188 patients. The 2D and 3D image information was integrated through channel stacking to create a fused image dataset. Three models were constructed using ResNet50 as the backbone network: 1) a 2D image model using the original dual-screen 2D images as input; 2) a 3D image model using reconstructed 3D images of the levator hiatus as input; and 3) a fused image model using fused 2D and 3D images as input. The diagnostic performance of each model was evaluated in the test set using accuracy, precision, recall, F1 score, and the area under the receiver operating characteristic curve (AUC). Differences in accuracy among the models were compared using the McNemar test. The 2D image model achieved an overall accuracy of 75.5%, with macro-averaged precision, recall, F1 score, and AUC values of 69.4%, 69.8%, 69.2%, and 0.878, respectively. The 3D image model achieved an overall accuracy of 76.1%, with macro-averaged precision, recall, F1 score, and AUC values of 78.7%, 71.7%, 71.3%, and 0.908, respectively. The fused image model achieved an overall accuracy of 82.3%, with macro-averaged precision, recall, F1 score, and AUC values of 80.5%, 78.7%, 79.3%, and 0.930, respectively. Its F1 scores for the 3 categories of no cystocele, mild cystocele, and significant cystocele were 80.4%, 68.4%, and 89.0%, respectively. The accuracy of the fused image model was significantly higher than that of the 2D image model (82.3% vs 75.5%, <i>P</i>=0.031). However, the difference between the fused image model and the 3D image model was not statistically significant (82.3% vs 76.1%, <i>P</i>=0.082). Deep learning models based on ultrasound images enabled automated diagnosis and grading of cystocele. The 2D and 3D ultrasound images each have distinct and complementary advantages in the assessment of cystocele, and their fusion may improve diagnostic accuracy and balanced performance across categories. The fused image model demonstrated significantly higher diagnostic accuracy than the 2D image model and showed potential for automated diagnosis and grading of cystocele.

Tags

Topics