Augmented-Reality Neuro-Navigation with AI Image Interpretation to Improve Accuray for Transcranial Magnetic Stimulation

The combined use of augmented-reality neuro navigation and AI scan interpretation can be used to improve the accuracy of transcranial magnetic stimulation in delivering treatment to recognised targets within the brain. With the advancement of AI interpretation of medical imaging, automatically identification of specific anatomical targets can be performed quickly and accurately. In patients with standard (non-functional) cranial imaging already performed, identifying specific targets would enable the individualised placement of the treatment coil. In patients without pre-treatment imaging, the use of a pre-defined model cortex, developed from the AI model with identified targets and morphed during registration to match the patient’s head, could still be used to improve on existing methods.

Different methods of TMS coil positioning have been documented¹⁺². The conventional "function-guided procedure" first locates the M1 reference point – representing the position over the primary motor cortex for the contralateral hand – and uses this to locate the pre-frontal cortex 5cm anterior to this in the sagittal place. These references are empiric, and do not consider the large variability in brain morphology³. The validity of this method has been investigated, with only 32% of subjects⁴ having their target dorsolateral pre-frontal cortex at this point. When looking at other potential TMS treatment targets for pathologies such as stroke and Parkinson's Disease, as the disease process impacts on the brain structure itself using a pre-defined one-size-fits-all approach becomes more problematic.

For the purposes of neuroradiology and neuro-navigation, AI-driven image interpretation and segmentation from standard MR imaging has become commonplace. Similar technology can be harnessed for use in delivering TMS. Firstly, a dataset of standard brain MRIs is required, and these are readily available commercially. Next, these need to be correctly annotated, with the common target regions mapped out by a neuro-radiologist. The images and target region mappings are then used to train and subsequently test a neural network. Once a proven algorithm has been developed it can be hosted on the cloud as an app service such as Microsoft's Azure. Using by an API, it can accept uploads of patient's standard DICOM MRI files, process them to define 3-dimensional meshes of target regions and the patient’s skin, and return them to a calling application on a treatment system.

A suggested workflow for a patient who already has standard MR imaging, for TMS therapy to the prefrontal dorsolateral cortex would be as follows:
A patient’s imaging would be uploaded in standard DICOM format to have target regions identified modelled in 3-dimensions, a skin mesh generated, and registration landmarks.
The navigation software on the treatment machine would download the patient’s individual plan and allow selection of the required treatment targets.
The navigation software would be used to register the position of the patient either with anatomical registration landmarks as is done currently, or using skin surface matching (as proven to deliver superior accuracy⁵).
The navigation software would display the real-time coil position with respect to the intended treatment target (for example, implemented in Unity3D). If also available to be viewed by the patient where appropriate, it could improve engagement and compliance with the treatment.
If patient's have undergone functional MR imaging and segmentation of target regions using 3rd party software with open file formats such as BrainLab, these could be uploaded and converted for use in the same way.

For patients without pre-treatment imaging, significant benefit can still be gained by the adoption of a standard plan of target regions, generated by averaging the training dataset used in the AI algorithm. This standard plan would replace the download of the patient’s individual plan in step 2 above, and still facilitate the landmark or surface matching registration, morphing the skin and target region meshes to match the patient anatomy.

With more accuracy in locating the relevant target region, more useful data can be collected from treatments to inform research and future practice. The overall aim is to deliver treatment more accurately to the required region to achieve better patient outcomes and safety, with minimal impact on the current proven workflow.

1. Herwig U, Schönfeldt-Lecuona C, Wunderlich AP, von Tiesenhausen C, Thielscher A, Walter H, et al. The navigation of transcranial magnetic stimulation. Psychiatry Res 2001;108:123—31.

2. Sparing R, Buelte D, Meister IG, Paus T, Fink GR. Transcranial magnetic stimulation and the challenge of coil placement: a comparison of conventional and stereotaxic neuronavigational strategies. Hum Brain Mapp 2008;29:82—96.

3. J.-P. Lefaucheur, Why image-guided navigation becomes essential in the practice of transcranial magnetic stimulation, Neurophysiologie Clinique/Clinical Neurophysiology, 2010;40;1;1-5.

4. Herwig U, Padberg F, Unger J, Spitzer M, Schonfeldt-Lecuona C. Transcranial magnetic stimulation in therapy studies: examination of the reliability of ‘‘standard’’ coil positioning by neuronavigation. Biol Psychiatry 2001;50:58-61.

5. Aino E Nieminen, Jaakko O Nieminen7, Matti Stenroos, Pavel Novikov, Maria Nazarova, Selja Vaalto, Vadim Nikulin and Risto J Ilmoniemi: Accuracy and precision of navigated transcranial magnetic stimulation. Journal of Neural Engineering 2022;19;6.