Predicting VIM Target Coordinates from a t1w MRI Image Using Machine Learning Models

Introduction: The ventral intermediate nucleus (VIM) of the thalamus is a key target for deep brain stimulation (DBS) and stereotactic lesioning—techniques including radiofrequency ablation, gamma knife radiosurgery, and recently MRI-guided focused ultrasound (MRgFUS), particularly for treating essential tremor. Traditionally, VIM targeting has relied on indirect coordinates derived from stereotactic atlases, but the precision of this atlas-based approach remains uncertain, with no consensus on optimal indirect targeting coordinates. While advanced imaging techniques can help delineate the VIM’s borders, access to these resources is often limited in resource-constrained neurosurgical centers.

Methods: We developed and evaluated machine learning (ML) models to predict stereotactic VIM coordinates using eight input features—coordinates of readily identifiable anatomical points around the thalamus from the axial AC-PC aligned T1-weighted MRI image from 100 healthy subjects of the Human Connectome Project (HCP). Four models—Linear Regression (LR), K-Nearest Neighbor (KNN), Support Vector Regression (SVR), and Deep Neural Networks (DNN)—were trained and optimized on 75 MRIs, and then tested on 25 MRIs. Error was measured as the distance between each model’s predicted coordinates and VIM centroids derived from THOMAS, a multi-atlas segmentation method based on white-matter-nulled MP-RAGE imaging.

Results: From the 25 unseen MRIs, the mean Euclidean error for LR was 1.34 ± 0.60 mm, KNN 1.23 ± 0.53 mm, SVR 1.36 ± 0.64 mm, and DNN 0.96 ± 0.42 mm. For DNN, the mean errors in the x, y, and z axes were 0.52, 0.64, and 0.56 mm, respectively.

Conclusion : Our findings suggest that ML models, particularly DNNs, can accurately predict VIM coordinates, potentially making VIM targeting more accurate and accessible in resource-limited settings, particularly in low- and middle-income countries (LMICs).