TY - JOUR
T1 - Automatic measurement of head circumference in fetal ultrasound images using a squeeze atrous pooling UNet
AU - Hekal, Asmaa A.
AU - Amer, Hanan M.
AU - Moustafa, Hossam El Din
AU - Elnakib, Ahmed
N1 - Publisher Copyright:
© 2024
PY - 2025/5
Y1 - 2025/5
N2 - Recently, artificial intelligence plays a significant role in automatically measuring anatomical structures from ultrasonic images. This paper presents an automated deep-learning segmentation of fetal heads from ultrasound images. In addition, the fetal head circumference (HC) is automatically determined. Initially, the fetal head is detected using a Yolov5 model and transfer learning. Data augmentation is applied to expand the deep-learning training set for fetal head segmentation. The fetal head is segmented using a modified UNet model, called a Squeeze Atrous Pooling (SAP)-UNet, which incorporates an Atrous spatial pyramid Pooling (ASPP) block and a squeeze and excitation (SE) block. Finally, an ellipse fitting is applied to measure the HC biometric. Experimental results on the online dataset HC18 achieve a Dice similarity coefficient (DSC) of 98.02% ± 0.8, Hausdorff distance (HD) of 1.20 ± 0.68 mm, HC difference (DF) of −0.08 ± 1.2 mm, and HC absolute difference (ADF) of 1.11 ± 1.3 mm. Experiments show that incorporating both the SE and ASPP blocks into the UNet model has successfully improved the segmentation performance. In comparison to the recent techniques for fetal head segmentation and HC biometry, the proposed SAP-UNet achieves superior performance. These results highlight the promise of using the proposed SAP-UNet model for accurate fetal head segmentation.
AB - Recently, artificial intelligence plays a significant role in automatically measuring anatomical structures from ultrasonic images. This paper presents an automated deep-learning segmentation of fetal heads from ultrasound images. In addition, the fetal head circumference (HC) is automatically determined. Initially, the fetal head is detected using a Yolov5 model and transfer learning. Data augmentation is applied to expand the deep-learning training set for fetal head segmentation. The fetal head is segmented using a modified UNet model, called a Squeeze Atrous Pooling (SAP)-UNet, which incorporates an Atrous spatial pyramid Pooling (ASPP) block and a squeeze and excitation (SE) block. Finally, an ellipse fitting is applied to measure the HC biometric. Experimental results on the online dataset HC18 achieve a Dice similarity coefficient (DSC) of 98.02% ± 0.8, Hausdorff distance (HD) of 1.20 ± 0.68 mm, HC difference (DF) of −0.08 ± 1.2 mm, and HC absolute difference (ADF) of 1.11 ± 1.3 mm. Experiments show that incorporating both the SE and ASPP blocks into the UNet model has successfully improved the segmentation performance. In comparison to the recent techniques for fetal head segmentation and HC biometry, the proposed SAP-UNet achieves superior performance. These results highlight the promise of using the proposed SAP-UNet model for accurate fetal head segmentation.
UR - http://www.scopus.com/inward/record.url?scp=85213218547&partnerID=8YFLogxK
UR - http://www.scopus.com/inward/citedby.url?scp=85213218547&partnerID=8YFLogxK
U2 - 10.1016/j.bspc.2024.107434
DO - 10.1016/j.bspc.2024.107434
M3 - Article
AN - SCOPUS:85213218547
SN - 1746-8094
VL - 103
JO - Biomedical Signal Processing and Control
JF - Biomedical Signal Processing and Control
M1 - 107434
ER -