Purpose or Case Report: To develop an AI-based automatic tool for Amniotic fluid volume (AFV) and fetal weight (FW) quantification that easily integrates into everyday diagnostic workflow
Methods & Materials: This single-center retrospective study at a large academic referral center reviewed electronic health records for all fetal MRIs performed from 01/2018 to 01/2022. A random sample of 109 fetal MRIs (gestational ages of 16-37 weeks) were analyzed, with scans performed on 3T or 1.5T Siemens magnets. All cases had both fetal MR and US within 1 week of each other. An innovative deep-learning architecture was explicitly designed for analyzing fetal MRI scans. An easily accessible web interface using Streamlit, an open-source Python framework, enabled rapid deployment of our medical imaging application. The backend leverages MRIcroGL's visualization capabilities to process and analyze the uploaded MRI scans. When users submit their fetal MRI data through the Streamlit interface, the system automatically performs the necessary segmentations using our trained deep-learning models and utilizes MRIcroGL's rendering engine to generate detailed 3D visualizations. This integration allows any radiologist with access to internet to obtain high-quality AF volume measurements with 3D reconstructions directly through their web browser without requiring specialized software installation. The platform's architecture ensures efficient processing while maintaining the security and integrity of sensitive medical data.
Results: Preliminary findings indicate that the AI-driven model outperforms existing AFV and FW manual estimation methods. Our automated segmentation model demonstrated excellent performance, achieving a Dice similarity coefficient of 0.961 compared to expert manual segmentations. Our model achieved a high accuracy of 0.96 for AFV classification by utilizing both amniotic fluid volume and fetal volume as key features. The classification model's robustness is further evidenced by its area under the curve score of 0.99, indicating discriminative solid ability in different AFV categories.
Conclusions: The AI model discussed allows calculation of AFV and FW by means of a highly accurate system publicly available. This is of particular interest when a correlating US is not available when reading a fetal MR. Its ease of use ensures that it can be widely adopted in clinical settings, providing real-time, accurate results via a simple online interface accessible to pediatric radiologists.