Conditional deep generative normative modeling for structural and developmental anomaly detection in the fetal brain

Abstract

Fetal brain development is a complex and dynamic process, and its disruption can lead to significant neurological disorders. Early detection of brain aberrations during pregnancy is critical for optimizing postnatal medical intervention. We propose a deep generative anomaly detection framework, conditional cyclic variational autoencoding generative adversarial network (CCVAEGAN), that can identify structural brain anomalies using fetal brain magnetic resonance imaging. CCVAEGAN leverages covariate conditioning on gestational age and cyclic consistency training to generate high-fidelity normative fetal brain images to enhance anomaly detection across various neurodevelopmental stages and diagnoses. Using MRI data from typically developing and clinically abnormal fetuses across multiple sites, CCVAEGAN achieves superior image generation quality and anomaly detection accuracy than other comparable models, consistently producing anatomically precise images with lower reconstruction errors and higher structural similarities. Anomaly detection experiments yielded near-perfect AUROC values (>0.99) across various anomaly score metrics, and visual assessments confirmed the model's ability to localize and characterize structural abnormalities. Additionally, external validation on separated-site cohorts demonstrated the generalizability of the CCVAEGAN framework, showing robust detection performance despite data variations. These findings demonstrate CCVAEGAN's potential as a powerful tool for automated, objective anomaly screening, that could significantly enhance the efficiency of clinical workflows for early diagnosis of fetal brain anomalies. Furthermore, this approach has the potential universality to apply to other medical imaging not limited to specific organs or imaging modalities in the future.