<p>Background The early and accurate identification of fetal cerebellar hypoplasia (CH) during the prenatal stage is crucial for timely intervention and decision-making. Medical ultrasound represents the primary tool for CH diagnosis. However, the accuracy of CH diagnosis may be limited by imaging artifacts and subjective judgments among sonographers. Artificial intelligence provides an effective tool for improving the diagnostic accuracy and consistency of ultrasound imaging. Objective This study aims to develop and validate a self-supervised learning radiomics nomogram (SSRN) that integrates the anatomical structures of the fetal skull, cerebellum, and cistern in order to assess prenatal risk for fetal CH, to identify significant factors that may influence CH diagnosis, and to evaluate the diagnostic efficacy of SSRN in clinical applications. Method This retrospective study included clinical data and ultrasound images from 547 normal fetuses and 301 fetuses diagnosed with CH between September 2019 and September 2023 at the Ultrasound Diagnostic Department of Hubei Maternal and Child Health Hospital, China. Subsequently, the standard brain views were selected by experienced sonographers, who also delineated the contours of the skull, cerebellum, and cistern in each view. In the self-supervised learning strategy, VQ-VAE-2 was employed to extract the latent features from these brain views, which were then converted into a self-supervised image score (SIS). Radiomics features were extracted from the combined region of interest (ROI) of the cerebellum and cistern to obtain a radiomics score (RS). The proposed SSRN was constructed by integrating significant demographic and morphological features identified through univariate and multivariate logistic regression analyses, along with SIS and RS. In order to validate the clinical potential of SSRN, several comparative models were established, including an expert model (EM) comprising three sonographers with varying years of clinical experience, a clinical regression model (CLM) based on clinical data, a self-supervised learning classification model (SSL-CM) based on latent features extracted by VQ-VAE-2, and a radiomics model (RM) based on radiomics features. Results The study identified several statistically significant influencing factors, including the width of the cistern, the area of the cistern, cerebellum and skull, the area ratio between the cistern and cerebellum, SIS, and RS. Subsequently, SSRN was constructed using the aforementioned factors, achieving an accuracy of 0.906 and an AUC of 0.956. This resulted in a significantly enhanced performance in comparison to alternative models and EM (accuracy: 0.782, AUC: 0.752). Furthermore, it was observed that the integration of anatomical structures in SSRN and CLM (accuracy: 0.894, AUC: 0.934) performed better than RM (accuracy: 0.871, AUC: 0.934) and SSL-CM (accuracy: 0.622, AUC: 0.641), which lacked such incorporation. A subgroup analysis revealed that the AUCs were significantly improved in subgroups with maternal age ≥ 30 years, gestational week ≥ 28.4 weeks, and above-average anatomical structural features. Conclusion SSRN demonstrated superior performance in identifying prenatal CH risk, highlighting the innovative application of self-supervised learning and anatomical features in prenatal CH screening. Furthermore, SSRN exhibited the potential to become a valuable supplementary tool for identifying CH in high-risk pregnancies, thereby promoting personalized prenatal management.</p>