Optimizing a deep learning model for the prediction of electric field induced by transcranial magnetic stimulation for mild to moderate traumatic brain injury patients

Abstract

Transcranial magnetic stimulation (TMS) is a non-invasive method for treating neurological and psychiatric disorders. It is being tested as an experimental treatment for patients with mild to moderate traumatic brain injuries (mTBI). Due to the complex, heterogeneous composition of the brain, it is difficult to determine if targeted brain regions receive the correct amount of electric field (E-field) induced by the TMS coil. E-field distributions can be calculated by running time-consuming finite element analysis (FEA) simulations of TMS on patient head models. Using machine learning, the E-field can be predicted in real-time. Our prior work used a Deep Convolutional Neural Network (DCNN) to predict the E-field in healthy patients. This study applies the same DCNN to mTBI patients and investigates how model depth and color space of E-field images affect model performance. Nine DCNNs were created using combinations of 3, 4, or 5 encoder and decoder blocks with the color spaces RGB, LAB, and YCbCr. As depth increased, training and testing peak signal-to-noise ratios (PSNR) increased and mean squared errors (MSE) decreased. The depth 5 YCbCr model had the highest training and testing PSNRs of 34.77 and 29.08 dB and lowest training and testing MSEs of 3.335∗10−4 and 1.237∗10−3 respectively. Compared to the model in our prior work, models of depth 5 have higher testing PSNRs and lower MSEs and, except for RGB. Thus, DCNNs with depth 5 and alternative color spaces, despite losing information through color space conversions, resulted in higher PSNRs and lower MSEs.