Real-time AI-guided planning of fetal femur bone MRI and automatic quantification of growth

Abstract While an important indicator of antenatal growth and routinely assessed by ultrasound, MRI-based femur length measurement remains challenging because of bone-slice misalignment, fetal motion, and manual assessment. We developed and evaluated a fully automated real-time method for acquisition planning and fetal femur length measurement. A low-latency 3D U-Net with transformer-based multi-representation refinement was trained with fractional optimal-transport supervision on 59 fetal MRI scans acquired at 0.55T (18-40 weeks) to localize the proximal and distal femur endpoints. These coordinates were then used to adapt a 4-second echo-planar imaging sequence in real time for in-plane femur coverage and length extraction. Retrospective evaluation was performed on 138 scans (0.55T and 1.5T, 19-39 weeks, including 19 pathological cases), and real-time testing was conducted in 24 cases (17-39 weeks). Across 129/138 cases, the method achieved a mean endpoint localization error of 5.1 mm and a femur length deviation of 3.3 mm (5.1 $$\%$$ ) relative to expert annotations and, in 57 cases, to matched clinical ultrasound measurements. Robust performance was maintained under domain shift, and internal consistency was demonstrated by a mean left-right length difference of 2.8 mm (4.4 $$\%$$ ) in 88 cases with bilateral extraction. Real-time assessment succeeded in 22/24 cases, supporting operator-independent fetal growth assessment and future large-scale evaluation in pregnancies with compromised growth.