Abstract
<title>Abstract</title> <p>Fetal brain MRI plays a crucial role in prenatal diagnosis, yet inevitable fetal motion and slice-to-volume reconstruction (SVR) errors introduce corruptions including noise and artifacts that degrade diagnostic reliability. Existing image enhancement methods either require noise-free references unavailable in fetal MRI or assume pixel-independent noise that mismatches the spatially coarse noise in this modality. Here, we introduce FetMRIE, the first unsupervised fetal brain MRI enhancement framework, based on adaptive state-matching denoising diffusion. Its two-stage architecture combines variational autoencoder-based initial denoising with diffusion-based structural restoration, balancing noise suppression and anatomical fidelity. Experiments on 2,481 fetuses across eight multi-center datasets, spanning four T2-weighted sequences (ss-FSE, TSE-SSH, HASTE, B-TFE), 1.5-T and 3.0-T scanners, and both normative and pathological cases (ventriculomegaly and germinal matrix-intraventricular hemorrhage), demonstrate state-of-the-art performance on unsupervised PSNR (uPSNR), unsupervised MSE (uMSE), tissue contrast t-score (TCT), and cross-tissue noise consistency (CNC) with robust cross-domain generalizability. Four radiologists with strong inter-rater agreement (correlation coefficients: 0.720–0.871) consistently rated FetMRIE highest in noise suppression, anatomical fidelity, and overall image quality. Three downstream tasks (gestational age prediction, anomaly detection, and biometry measurement) further demonstrate that FetMRIE improves the accuracy and reliability of automated clinical analyses by enhancing normative anatomical clarity while faithfully preserving pathological signatures.</p>