Atrous spatial pyramid pooling assisted automatic segmentation model and ellipse fitting approach based fetal head segmentation and head circumference measurement

Abstract

Fetal head circumference (HC) is an important biometric measurement that is useful in obstetric clinical practice to assess fetal development. Existing methods for fetal head circumference measurement have limitations in accurately capturing the shape of the fetal skull, leading to potential errors in clinical assessments. In this study, the Atrous spatial pyramid pooling assisted multi-scale feature aggregation automatic Segmentation (ASPPA-MSFAAS) model is introduced. The ASPPA-MSFAAS model addresses these limitations by incorporating multi-scale feature extraction and aggregation, enabling more precise segmentation and measurement of the fetal head. The objective of the multi-scale segmentation model is to improve fine-grained HC measurement and segmentation performance by learning multiple features under different sensitivity fields. Initially, pre-processing stages are applied to input images in order to eliminate undesired distortions. The ASPPA-MSFAAS model contains three modules: Atrous spatial pyramid pooling multi-scale feature extraction module (ASPP-MSFEM), multi-scale feature aggregation module (MSFAM), and Attention module. During the training and testing stages, these three modules are utilized to precisely segment the intricate location of the fetal head (FH). Post-processing operations are then used to smooth the region and eliminate extraneous artifacts from the segmentation results. Post-processing results are subjected to an ellipse fitting approach to get HC. Evaluation results show that the proposed approach attains 99.12 %±0.6 DSC and 99 %±1.99 MIoU using the HC 18 grand challenge dataset. Also, the proposed approach attained 98.99 % DSC, 1.287 HD, and 0.334 ASD performance for the Large-scale annotation dataset (National Library of Medicine).