ARANet: Adaptive Resolution Attention Network for Precise MRI-Based Segmentation and Quantification of Fetal Size and Amniotic Fluid Volume

Abstract

Amniotic fluid volume (AFV) is a critical indicator of fetal health, traditionally assessed using ultrasound-based methods, which are limited by operator dependency and 2D measurements. While MRI offers superior tissue characterization, it remains underutilized for AFV assessment due to labor-intensive manual segmentations. Pulse sequence variations in MRI can significantly influence image contrast through T1 and T2 weighting and potentially cause signal dropout in certain regions, making automated analysis challenging. To address these challenges, we present ARANet (Adaptive Resolution Attention Network), featuring a novel adaptive resolution attention module that uniquely combines adaptive resolution processing with channel-wise attention mechanisms for precise MRI-based segmentation. In extensive evaluations, ARANet demonstrates superior performance with a Dice score of 0.961 at full brightness, maintaining robust performance (0.909) even at 30% brightness, significantly outperforming existing models in challenging low-contrast conditions. We introduce FAFO3D, a new metric integrating amniotic fluid overlap, fetal age, and volume, which achieves 100% accuracy in clinical AFV classification (normal, oligohydramnios, and polyhydramnios) when combined with a support vector classifier. A comprehensive web platform automates the workflow from DICOM processing to analysis, providing detailed reports and 3D visualizations for clinical implementation. This solution offers a novel, accurate, and clinically applicable approach for AFV assessment, though further validation in diverse clinical environments is recommended.