Unsupervised spatial registration of fetal cranial ultrasound slices for radiomics analysis

Ultrasound examinations are clinically required for pregnant women to determine whether the fetal brain is developing normally. Doctors often need to manually adjust the 2-dimensional (2D) ultrasound probe to obtain a standard cranial plane for conducting fetal cranial ultrasound radiomics analysis, a process that is both challenging and time-consuming. To address this issue, this paper proposes an unsupervised registration method based on contrastive learning aimed at predicting the position of the 2D slice in 3-dimensional (3D) space. We exploit the anatomical consistency of adjacent slices in cranial space and introduce a novel pretext task. This task enables the contrastive learning model to learn the features of 2D slices at different spatial locations. Then, an automated spatial sampling method is used to construct a retrieval database that contains anatomical and spatial information for slices at various positions, which is used to perform spatial registration of 2D slices at arbitrary locations. We evaluated our method using 39 cases of mid-gestation 3D ultrasound data from the hospital. The experimental results demonstrate that our method achieves excellent performance across four registration metrics, with plane angle at 6.870°, Euclidean distance at 8.969 voxels, normalized cross-correlation at 0.882, and structural similarity at 0.771. The results confirm the efficiency of our method with small sample data and demonstrate its reliable generalization and transferability. This method can provide clinicians with precise positional guidance for the rapid localization of standard planes, which can be used for subsequent fetal cranial radiomics research. Moreover, our approach does not require predefined registration standards or manual annotations, and it can be extended to other registration tasks, offering a reference for researchers.