A Transfer Learning Approach to Localising a Deep Brain Stimulation Target

The ventral intermediate nucleus of thalamus (Vim) is a well-established surgical target in magnetic resonance-guided (MR-guided) surgery for the treatment of tremor. As the structure is not identifiable from conventional MR sequences, targeting the Vim has predominantly relied on standardised Vim atlases and thus fails to model individual anatomical variability. To overcome this limitation, recent studies define the Vim using its white matter connectivity with both primary motor cortex and dentate nucleus, estimated via tractography. Although successful in accounting for individual variability, these connectivity-based methods are sensitive to variations in image acquisition and processing, and require high-quality diffusion imaging techniques which are often not available in clinical contexts. Here we propose a novel transfer learning approach to accurately target the Vim particularly on clinical-quality data. The approach transfers anatomical information from publicly available high-quality datasets to a wide range of white matter connectivity features in low-quality data, to augment inference on the Vim. We demonstrate that the approach can robustly and reliably identify the Vim despite compromised data quality, and is generalisable to different datasets, outperforming previous surgical targeting methods.