Optimized fetal head circumference estimation in 2D ultrasound using EfficientNet-B7 and Adam optimizer.

Accurate measurement of fetal head circumference (HC) is critical for monitoring fetal growth and the diagnosis of microcephaly and macrocephaly. However, conventional manual 2D ultrasound measurements tend to vary because of factors such as the presence of amniotic fluid, acoustic shadowing, occlusions, and variations in fetal position. To address these problems, this study proposes a regression-based deep learning model using EfficientNet-B7 and the Adam optimizer to enhance the accuracy and robustness of HC measurement. The proposed model utilizes EfficientNet’s powerful feature extraction capabilities and compound scaling, along with Mean Squared Error loss to minimize prediction errors. The model was trained on the HC-18 training set and evaluated on a test set consisting of 335 images, the proposed approach yielded a normalized MAE of 0.028 and an HC-specific MAE of 1.978 mm. Although this error rate is higher than that of segmentation-based approaches on the entire HC-18 dataset (usually 1–2 mm), the regression model shows the feasibility of fast, fully automatic HC measurement with comparatively low model complexity and an inference time of approximately 0.03 s. These results demonstrate the trade-off between geometric modeling accuracy and computational efficiency. Future work will focus on training the model on larger and more diverse datasets, accelerating processing, and investigating alternative training methods to better address the challenges of ultrasound imaging.