OSAM-NET: A multi-feature fusion model for measuring fetal head flexion during labor with transformer multi-head self-attention

Fetal head flexion is essential during labor. The current assessment presents technical challenges for unskilled ultrasound operators. Therefore, the study aimed to propose an occiput-spine angle measurement network (OSAM-NET) to improve the accuracy and intrapartum applicability of fetal head flexion assessment. We used YOLOv8 to extract key anatomical structures (occiput and spine) and Vision Transformer to extract global features on ultrasound images of the posterior mid-sagittal section of the fetus (PMSF). Then, by fusing the geometric location information and global features through the multi-head self-attention mechanism, we constructed the multi-feature fusion model OSAM-NET. The model was able to extract intricate features from multi-dimensional information, effectively boosting the accuracy of occiput-spine angle (OSA) prediction. We curated the first OSA dataset comprising 1688 high-quality, clear PMSF ultrasound images and annotations to train and test our model. We evaluated the performance of OSAM-NET and compared it with other models. The results showed that OSAM-NET outperformed the comparison models on all evaluation metrics, with R² increasing by nearly 13 %, and root mean square error (RMSE) and mean absolute error (MAE) decreasing by approximately 15 % and 20 %, respectively. The intraclass correlation coefficient (ICC) improved by about 8 %, with the average value reaching 0.89, indicating good agreement with the measurements of ultrasound experts. The multi-feature fusion model OSAM-NET demonstrates strong applicability and predictive accuracy for complex ultrasound images. This study provides a reliable and efficient tool for automatically evaluating fetal head flexion. The real-world application potential has been validated on prospective test dataset.