An image quality transfer approach for localising deep brain stimulation targets

Area of Science:

• Neurosurgery
• Medical Imaging
• Machine Learning

Background:

• The ventral intermediate nucleus (Vim) is a key target for tremor treatment in functional neurosurgery.
• Conventional MRI lacks contrast for precise Vim targeting, leading to reliance on atlases that ignore individual anatomy.
• Current connectivity-based Vim targeting methods require high-quality diffusion imaging, often unavailable clinically.

Purpose of the Study:

• To develop a novel transfer learning approach for accurate Vim targeting using clinical-quality MRI data.
• To overcome limitations of traditional atlases and connectivity-based methods in the presence of anatomical variability and low-quality imaging.
• To enable robust Vim identification despite compromised data quality and diverse clinical acquisition protocols.

Main Methods:

• A transfer learning technique was employed to map anatomical information from high-quality datasets to low-quality data.
• The method leverages white matter connectivity features to enhance Vim inference.
• The approach was validated on clinical-quality data with varying degrees of quality issues.

Main Results:

• The proposed method accurately and reliably identifies the Vim even with compromised diffusion imaging data quality.
• The approach demonstrated generalizability to unseen clinical data acquired with different protocols and severe quality issues.
• The transfer learning method successfully augmented inference on the Vim using low-quality data.

Conclusions:

• This novel transfer learning approach offers a robust solution for targeting the Vim in clinical settings.
• The method enhances the precision of functional neurosurgery for tremor treatment by accounting for individual variability.
• The approach is adaptable for targeting other deep brain structures beyond the Vim.