Subthalamic Nucleus and Substantia Nigra Automatic Segmentation Using Convolutional Segmentation Transformers (Conv-SeTr)

The Subthalamic Nucleus and Substantia Nigra have an important role in the treatment of Parkinson’s Disease (PD); however, they are difficult to identify on magnetic resonance imaging (MRI) and are of paramount importance in PD, which requires their precise localization. We present a pipeline methodology that allows autonomous segmentation of both structures, based on MRI T2-weighted images and Deep Learning techniques. Three segmentation architectures were compared: CLCI-Net, 3D U-Net and Conv-SeTr. All models were trained in two instances: the first with 60 T2-weighted standard protocol images from 1.5T MRI. Transfer learning was applied for the second instance in which the models were trained with 20 T2-weighted adjusted protocol images from 3T MRI. In all cases, the ground truth was obtained through manual segmentation by experts. All models produced an image mask with the segmented labelled structures co-registered to native space. The proposed transformer-based model segmented the volumes of interest with a DICE coefficient of 0.81 and an AVD of 0.06, which outperformed the other architectures. The Conv-SeTr presented promising results for segmenting Subthalamic Nucleus and Substantia Nigra, key structures in Parkinson’s disease research and treatment.