Neural network-based segmentation of dentate nuclei from 3T b0 brain images: a semi-supervised approach

The dentate nuclei (DN) of the cerebellum are essential for communication with the brain and are part of the Dentate Rubro-Talamic (DRT) Tract, involved in movement disorders such as Essential Tremor (ET). Accurate DN segmentation is crucial for clinical and research applications. Traditional methods like manual delineation or atlas-based approaches are time-consuming and variable. Some use uncommon MRI sequences, like susceptibilityweighted imaging. In young adults, B0 images at 3T offer a more accessible option. In this study, we implemented a semi-supervised method to segment DN on 3T b0 DWI images, which are widely available. A fully Convolutional Neural Network was developed using T1 and b0 images. HCP dataset images were used to generate binary masks for training. Performance was compared to a model based on nnUNet. Preliminary results showed a Dice Score (DS) of 0.85 ± 0.05 for our model vs. 0.83 ± 0.08 for nnUNet, indicating high accuracy. Our method is promising for planning neurosurgical treatment of essential tremor.