Automated Fetal Brain Segmentation in Slice to Volume Reconstruction in Clinical Practice

PURPOSE: To evaluate the potential for automatically derived volumetric data sets in slice to volume reconstruction (SVR) MRI images of the fetal brain to replace manually acquired 2D brain measurements in clinical practice. MATERIALS AND METHODS: This retrospective, IRB approved study included pregnant women undergoing fetal MRI with SVR. In addition to performing the standard of care fetal MRI protocol, 2D T2-SSFSE images were also collected at 2.5 or 3 mm slice thickness with 1.25 mm or 1.5 mm overlap in up to 5 imaging planes (2 independent/separate acquisitions per imaging plane) and used as inputs for SVR. The SVR pipeline was implemented using the Philips Research Imaging Development Environment (PRIDE) and the resulting color-coded brain segmentation maps and associated volumetric measurements were made available automatically through a web-based platform. Segmentation maps were rated on a 3-point scale for accuracy by the study radiologists (0 = poor, 1 = fair, 2 = good). 2D measurements from the clinical radiology reports were extracted and compared with the automated volumetric measurements. RESULTS: SVR was performed in 70 patients undergoing clinical fetal MRI (28.2 ± 4.3 weeks gestational age). Total SVR reconstruction time, including volumetric analysis, was 8.6 ± 3.4 minutes, which positively correlated with gestational age (r = 0.78). Segmentation quality was judged to be good in the majority of patients (70%, 49/70). 17.1% (12/70) of patients were considered to have poor segmentation map quality. The mean gestational age of these fetuses was significantly lower (25.1 ± 4.1 weeks) than the remaining fetuses (28.9 ± 4.1, p=0.004). Multiple strong positive correlations were observed between the linear measurements and the volumetric measurements including, a) the supratentorial brain volume with the linear brain biparietal diameter (r=0.94) and fronto-occipital diameter (r=0.92), b) total cerebellar volume with the transverse cerebellar diameter (r=0.92) and c) the total brainstem volume with the anterior-posterior pons measurement (r=0.88). CONCLUSIONS: Automated fetal brain segmentation in SVR is successful in the majority of cases, and the associated generated volumetric measurements are strongly correlated with manual 2D measurements, suggesting that they may eventually serve as a practical alternative in clinical practice.