Automatic classification of Levator ani muscle avulsion in 3D transperineal ultrasound images

Abstract

Levator ani muscle (LAM) avulsion is a common traumatic injury of the pelvic floor muscle occurring during vaginal childbirth and is linked to the development of pelvic organ prolapse (POP). POP is a pelvic floor disorder that affects up to 40% of women during their lifetime. Pelvic floor ultrasound imaging is used to diagnose LAM avulsion, but it requires trained experts and is time-consuming, leading to weeks-long delays in receiving diagnostic results and treatment. The purpose of this study is to demonstrate the feasibility of a deep learning system to automatically classify the degree of LAM avulsion from 3D transperineal ultrasound images (TPUS). Methods 3D TPUS images of the pelvic floor from 150 patients with and without POP-related LAM avulsion were collected. Out of these, 113 patients were included in the study. Over 650 key slices were extracted from the ultrasound volumes and cropped to a region of interest. A two-stage cascading ensemble architecture was developed, combining three convolutional neural networks (MobileNetV3-Small, EfficientNet- B0, and RegNetY-800MF) with a final decision layer. The system performs hierarchical classification: first distinguishing between normal and avulsion cases, then determining unilateral versus bilateral involvement, and finally classifying the degree of avulsion. Results In 5-fold cross-validation, the ensemble model demonstrated strong performance in both binary classification tasks. For avulsion detection, it achieved 86% accuracy, 88% sensitivity, 85% specificity, and an AUC of 0.94, consistently outperforming individual base classifiers (which achieved AUCs of 0.71 to 0.75). For bilateral/unilateral classification, the model achieved 80% accuracy, 82% sensitivity, 78% specificity, and an AUC of 0.87. When evaluated on a test set for final patient-level classification across five classes (normal, complete bilateral avulsion, complete unilateral avulsion, partial bilateral avulsion, and partial unilateral avulsion), the system achieved 46% accuracy. Conclusion This study demonstrates the feasibility of a deep learning classification system to automatically classify the degree of LAM avulsion from 3D TPUS images. While the system showed promising performance in binary classification tasks, its sequential decision-making design means that a single slice misclassification in the first classification stage can impact the final patient-level accuracy. Additionally, the limited dataset size can hinder the model’s ability to generalize effectively to unseen cases. Despite these limitations, the developed system shows the potential to expedite LAM avulsion diagnosis, overcoming the time constraints of manual diagnosis. This approach can broaden screening access, benefiting areas with limited healthcare resources, by reducing expert reliance and enabling timely treatment. Future work with larger and more diverse datasets could help address current limitations and further improve classification accuracy.