Anatomy-guided Test-Time Adaptation for Automated Fetal Brain MRI Morphometry

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

Keywords: Fetal brain MRI, Landmark detection, Test-time adaption, Morphometry

TL;DR: We propose an anatomy-guided test-time adaptation method integrating a local-global dual-network for cross-domain fetal brain MRI morphometry.

Abstract: Fetal brain MRI enables prenatal diagnosis of neurodegenerative diseases through linear morphologic measurements. Traditional manual measurements derived from the visual assessment of 2D MRI slices is labor-intensive, expertise-dependent, and suffers from high inter- and intra-rater variability due to inconsistent slice selection. Deep learning-based automated fetal brain MRI morphometry has been proposed to address these limitations. However, these automated models still struggle with limited generalizability due to cross-device MRI variability and lesion heterogeneity. To solve the problem, we propose an anatomy-guided test-time adaptation (TTA) method integrating a local-global dual-network with anatomical priors via atlas registration, enhancing cross-domain adaptability. Experimental results demonstrate that the model with TTA achieves superior brain measurement accuracy, outperforming both the model without TTA and the registration-based method.

Submission Number: 126