Deep Learning-Enabled Automated B-mode Segmentation for Regional Cerebral Perfusion Assessment in Neonates Using Ultrafast Power Doppler.

Cerebral perfusion disturbances contribute to neonatal brain injury, and ultrafast power Doppler can quantify regional cerebral blood volume but relies on manual segmentation that limits clinical use. We developed and validated a U-Net model for automated regional brain segmentation on B-mode images derived from neonatal transfontanellar ultrafast ultrasound to enable rapid arterial and venous cerebral blood volume quantification. Time-integrated, log-compressed ultrafast B-mode images from 80 neonates were manually segmented by an expert into the cortico-sub-cortical areas, the cingulate gyrus and the basal ganglia, and then used to train a U-Net with class-weighted loss (a 70/15/15 train/validation/test split). The trained model was then applied to segment B-mode images, with the resulting regions of interest used to quantify arterial and venous cerebral blood volume from the corresponding power Doppler data. Segmentation performance was quantified using the Dice similarity coefficient, with agreement between artificial intelligence- and manually derived cerebral blood volume assessed by Bland-Altman analysis. The median Dice similarity coefficients were 83%, 84% and 88% for the three regions. Mean Bland-Altman differences were ≤0.103 dB (arterial) and ≤0.096 dB (venous). Perturbation of the predicted masks by ±35% surface area change showed the narrowest limits of agreement at the unperturbed prediction, changing by no more than 0.1 dB between -5% and +5%, demonstrating automated segmentation as a rapid, reproducible alternative to manual delineation.