Automated Segmentation of Midbrain Structures in High-Resolution Susceptibility Maps Based on Convolutional Neural Network and Transfer Learning.

Background: Accurate delineation of the midbrain nuclei, the red nucleus (RN), substantia nigra (SN) and subthalamic nucleus (STN), is important in neuroimaging studies of neurodegenerative and other diseases. This study aims to segment midbrain structures in high-resolution susceptibility maps using a method based on a convolutional neural network (CNN).

Methods: The susceptibility maps of 75 subjects were acquired with a voxel size of 0.83 × 0.83 × 0.80 mm on a 3T MRI system to distinguish the RN, SN, and STN. A deeply supervised attention U-net was pre-trained with a dataset of 100 subjects containing susceptibility maps with a voxel size of 0.63 × 0.63 × 2.00 mm to provide initial weights for the target network. Five-fold cross-validation over the training cohort was used for all the models' training and selection. The same test cohort was used for the final evaluation of all the models. Dice coefficients were used to assess spatial overlap agreement between manual delineations (ground truth) and automated segmentation. Volume and magnetic susceptibility values in the nuclei extracted with automated CNN delineation were compared to those extracted by manual tracing. Consistencies of volume and magnetic susceptibility values by different extraction strategies were assessed by Pearson correlation coefficients and Bland-Altman analyses.

Results: The automated CNN segmentation method achieved mean Dice scores of 0.903, 0.864, and 0.777 for the RN, SN, and STN, respectively. There were no significant differences between the achieved Dice scores and the inter-rater Dice scores ( > 0.05 for each nucleus). The overall volume and magnetic susceptibility values of the nuclei extracted by the automatic CNN method were significantly correlated with those by manual delineation ( < 0.01).

Conclusion: Midbrain structures can be precisely segmented in high-resolution susceptibility maps using a CNN-based method.