FHP-ClsNet: a model for fetal head position classification based on intrapartum ultrasound.

Objectives: This study aims to develop an advanced deep learning model for accurate classification of fetal head positions (occiput anterior [OA], occiput posterior [OP], and occiput transverse [OT]) in intrapartum ultrasound images, which is crucial for labor management and delivery outcome prediction. Methods: We proposed FHP-ClsNet, a model integrating CBAM, ECA, and PSA attention mechanisms with an AIFI feature interaction module, based on YOLOv8 architecture. The study utilized 69 annotated intrapartum ultrasound videos divided into training, validation, and test sets. Data augmentation strategies including random flip, geometric transformation, and Mosaic enhancement were applied. Model performance was evaluated using accuracy (ACC), precision-recall area under curve (PR-AUC), receiver operating characteristic area under curve (ROC-AUC), and kappa coefficient. Results: The model achieved exceptional performance metrics: 0.8961 ACC, 0.8899 PR-AUC, 0.9401 ROC-AUC, and 0.8434 kappa score. Ablation experiments confirmed the significant contributions of attention mechanisms and feature interaction modules. Compared to baseline models, our model improved ACC by 1.6% and kappa by 3%. In video classification, it correctly identified 86/109 cases, outperforming other models. Conclusion: FHP-ClsNet demonstrates high accuracy in fetal head position classification from static ultrasound images, showing potential to assist clinical decision-making in labor management, though further improvements are needed for video sequence processing.