The main aim of STEM-DEEP (Stochastic electromagnetic modeling and deep learning for an effective and personalized transcranial magnetic stimulation) is to develop an innovative method for the identification of the optimal intensity of the electric current and coil position when using Transcranial Magnetic Stimulation (TMS) to stimulate a given cortical brain region in individual subjects.
To this aim, a cutting-edge approach is proposed. personalized data from Magnetic Resonance Imaging (MRI) images will be used and the variability of tissue properties (in terms of uncertainty quantification) will be considered. The forward field problem, i.e., to find the electric field distribution given the geometry, the tissue properties and the coil current, will be solved by applying a Model Order Reduction (MOR) technique; finally, the inverse problem, i.e., to find the coil current and coil position in order to obtain an effective TMS, will be solved in an accurate and innovative way thanks to Deep Learning (DL) techniques. The method will be validated by means of TMS measurements.
Twenty-five healthy subjects will be enrolled: starting from MRI images, a Finite Element Model (FEM) will be implemented. Based on the FEM, a MOR technique coupled with Polynomial Chaos Expansion (PCE) will provide a fast surrogate model that takes into account the variability of electrical conductivities within different brain tissues. This surrogate model will be used to create a database of electric field maps, firstly with regards to the brain region of the hand and subsequently with regards to the leg area.
An experimental session of TMS on the twenty-five subjects will provide the TMS intensity needed for effectively stimulating the hand area. The direct problem will be addressed by modeling these experiments with FEM with the aim of calculating the field distribution in the hand area that is necessary to obtain stimulation. Then the inverse problem of identifying the coil position, given the desired field distribution, will be dealt with. To this end, a Variational Autoencoder (VA) and a Convolutional Neural Network (CNN) will be trained. Finally, the TMS intensity will be scaled to properly match the field value identified from the field simulation based on the experiments. The same procedure will be repeated for the leg area, but neither other experiments nor field models will be needed. Finally, a validation session of experiments will be performed to assess our results.
To sum up, the proposed method aims at improving the diagnostic and therapeutic potential of TMS technique, with the main goal to avoid too stressful sessions, which could be even useless if a target area is not properly stimulated.
