A new approach to automatic measure fetal head circumference in ultrasound images using convolutional neural networks.

Fetal head circumference (HC) is an important biological index in prenatal ultrasound screening. In the clinic, fetal HC is usually measured manually by sonographers in two dimensional (2D) ultrasound images. The manual method is significantly affected by the inter/intra-observer difference and the process of manual measurement is inconvenient and time-consuming for sonographers. Although several artificial intelligence (AI) approaches had been applied to fetal HC measurement, they had weak generalization ability, especially for the incomplete or blurred skull edge. In this study, a fast and accurate method for fetal HC auto-measurement was proposed. Different from the common region segmentation method, an end-to-end convolutional neural network (CNN) for fetal skull boundary segmentation in 2D ultrasound images is proposed, which is an efficient method to directly segment the boundary of fetal skull by using the proposed double-branch structure. The segmentation results can be directly used to calculate fetal HC without complex post-processing. The proposed approach achieved excellent results: Mean Dice Sore (MDS)±std: 97.98 ± 1.30, Mean Hausdorff Distance (MHD)±std: 1.20 ± 0.68 mm, Mean Absolute Difference (MAD)±std: 1.75 ± 1.60 mm, Mean Difference (MD)±std: 0.08 ± 2.37 mm. Additionally, we drew a Bland-Altman plot to demonstrate that HC measured by the proposed approach has high agreement with the real value. Comprehensive results show that the proposed approach is comparable to the state-of-the-art methods for fetal HC measurement. Meanwhile, our approach belongs to a lightweight network with less parameters, which is convenient for deployment. We hope it could provide help for precision medicine in prenatal ultrasound screening.