Improving Basal Ganglia Segmentation in Neonatal Brain for Dilated Perivascular Space Assessment Using Soft Labels

Motivation: In neonates, dilated perivascular spaces are typically visible in the basal ganglia, where accurate boundary delineation of the basal ganglia is essential for reliable dilated perivascular space quantification. However, neonatal MRI data often exhibit low image quality and poor contrast, making precise boundary identification challenging. Goal(s): To develop a deep learning-based basal ganglia segmentation model for neonatal MR images. Approach: The model was trained using binarized (hard) labels and soft labels, and the results were compared. Results: The results demonstrate that the model using soft labels performs better in delineating the boundaries of the basal ganglia, particularly in the inferior slices. Impact: The proposed soft label-based segmentation method improves basal ganglia segmentation performance in low-contrast neonatal MR images compared to conventional hard label-based methods. The proposed method could be beneficial for perivascular space assessment in neonatal populations.