Automatic detection of the fetal brain midsagittal plane on MRI using a deep learning pipeline.
Authors
Affiliations (5)
Affiliations (5)
- EA fetus 7328 and LUMIERE Platform, University of Paris; School of Automation, Hangzhou Dianzi University, Hangzhou, China.
- EA fetus 7328 and LUMIERE Platform, University of Paris; Department of Radiology, Necker-Enfants Malades Hospital, APHP, Paris, France.
- School of Automation, Hangzhou Dianzi University, Hangzhou, China.
- EA fetus 7328 and LUMIERE Platform, University of Paris; Department of Obstetrics, Fetal Medicine and Surgery, Necker-Enfants Malades Hospital, APHP, Paris, France.
- EA fetus 7328 and LUMIERE Platform, University of Paris. Electronic address: [email protected].
Abstract
Accurate localization of the midsagittal plane (MSP) is essential for evaluating midline brain structures such as the corpus callosum and posterior fossa on fetal MRI. However, this task is often challenging due to fetal motion and the relatively large slice thickness of conventional acquisitions. Automated MSP detection could enhance image quality evaluation and promote a more reliable and standardized assessment of the fetal brain anatomy. To develop an automated deep learning pipeline for automatic fetal brain MSP detection from fetal brain MRI scans METHODS: : We retrospectively included fetuses with normal brain anatomy between 18 and 36 weeks who underwent routine 1.5 T fetal brain MRI using single-shot fast spin-echo T₂-weighted sequences. Ground-truth MSP were annotated by an expert in fetal MRI, and cases were labeled as MSP-present or MSP-absent. A multitask 2D U-Net was trained to identify the MSP by detecting the concurrent visibility of four midline structures: the genu and splenium of the corpus callosum, the cerebellar vermis, the pons, and to generate structure-specific probability maps and slice-level MSP scores, defined as the predicted probability that each sagittal slice corresponds to the true MSP. These were aggregated by a LightGBM classifier to produce a case-level decision. Model performance was assessed using five-fold cross-validation, with 80% of cases used for training and 20% for testing in each iteration. Evaluation metrics included area under curve (AUC), accuracy, sensitivity, specificity, and localization accuracy (distance between predicted and expert-annotated MSP). A total of 432 fetal brain MRI stacks from 225 fetuses were included at 32 [interquartile range, 28-34] weeks, of which 135 contained an identifiable MSP and 297 did not. The model achieved excellent slice-level performance to detect MSP, with an AUC of 0.95 and an accuracy of 0.87. At the case level, the AUC and accuracy were 0.81 and 0.74, respectively. In 99% of true-positive cases, the predicted MSP was localized within ±1 slice of the expert annotation, and in 75% of cases the correspondence was exact. The proposed deep learning framework accurately and consistently identifies the fetal brain MSP on routine MRI. Automated localization of the MSP offers an objective quality control measure, facilitating more accurate anatomical and biometric evaluations of the fetal brain.