Deep learning-driven automatic segmentation of uterine peristalsis in intracavitary electrohysterography signals

BACKGROUND AND OBJECTIVE: Uterine peristalsis, the rhythmic contractions of the inner layer of the myometrium during the menstrual cycle, is more challenging to identify than potent contractions during menstruation, pregnancy, or childbirth. Peristalsis plays a significant role in sperm ascension and embryo implantation, making its study valuable for fertility research. Intracavitary Electrohysterography (IC-EHG) is a promising technique for the electrophysiological assessment of uterine activity, but identifying basal, contraction, and artifact segments is a task currently performed by experts consuming substantial time, and is affected by the expert's subjectivity. This study aims to develop a deep learning model to aid clinicians in this segmentation task. METHODS: A total of 306 IC-EHG signals, amounting to a total duration of 9318 min, were collected from three different clinical centers. The model architecture, based on a modified version of the U-Time model, was evaluated using event-oriented evaluation performance metrics (accuracy, recall, precision, and F1-score), episode-oriented evaluation performance metrics (Margin Validation Test, including full detection, partial detection, false detection, non-detection, and others), and typical IC-EHG contraction parameters for signal characterization (root mean square amplitude, contraction frequency, and duration). RESULTS: Event-oriented evaluation performance metrics results indicate accurate classification of basal, contraction, and artifact classes (mean F1-score (%): 94.6, 92.3, and 92.3, respectively). Episode-oriented evaluation performance metric results underscore model's ability to detect consistent events (basal (full + partial detection:) 88.9 + 7.3%, contraction: 84.9 + 7.7%, and artifact: 75.3 + 16.8%). Uterine peristalsis parameters derived from IC-EHG contraction events exhibited low mean absolute errors between manual and model-based segmentations: contraction frequency (4.9%), root mean square (2.8%), and duration (6.8%), along with high agreement between both approaches (ICC ≥ 0.92). CONCLUSIONS: The findings underscore the model's reliability and robustness. The developed deep learning-based model offers significant clinical value by not only saving time in the challenging task of uterine peristalsis segmentation but also reducing expert variability. The developed automatic annotation system has the potential to serve as a valuable instrument in the context of infertility and associated pathologies.