Comprehensive Evaluation of Unsupervised Image Enhancement for Volumetric Fetal Brain MRI

Yingqi Hao, Mingxuan Liu, Hongjia Yang, Haoxiang Li, Xuguang Bai, Yi Liao, Haibo Qu, Qiyuan Tian

Keywords: Fetal brain MRI, Image enhancement, BME-X, Foundation model, 3D VAE.

TL;DR: Comprehensive evaluation of unsupervised methods for enhancing volumetric fetal brain MRI reveals that a pre-trained 3D VAE surpasses BME-X in segmentation, contrast, and lesion fidelity.

Abstract: MRI provides superior soft tissue contrast over ultrasound, making it essential for evaluating fetal brain development and pathology. However, clinical use of 2D thick-slice T2-weighted imaging remains constrained by motion-induced artifacts that degrade both image quality and subsequent quantitative analyses. While existing volumetric reconstruction pipelines (e.g., NiftyMIC, NeSVoR) incorporate motion correction, their outputs often retain residual noise due to a lack of post-reconstruction enhancement solutions. Although the recently proposed foundation model BME-X represents the first dedicated approach for fetal MRI enhancement, its generalizability to heterogeneous clinical datasets remains unproven. To bridge this gap, we conduct the first comprehensive comparison of BME-X and other unsupervised image enhancement methods on normal and pathological fetal brain MRI, based on tissue segmentation accuracy, tissue contrast t-score (TCT), lesion fidelity, and reader assessment. Results show that a pre-trained 3D convolutional variational autoencoder (VAE) achieves more effective enhancement compared to BME-X. Code and pre-trained weights are available at: https://github.com/yingqihao2022/FetalBrainEnhancement.

Submission Number: 11