Dissertations and Theses
Date of Award
2023
Document Type
Thesis
Department
Computer Science
First Advisor
A. Duke Shereen
Keywords
tDCS, MRI field mapping, machine learning
Abstract
Background: Transcranial direct current stimulation (tDCS) is a promising non-invasive method for treating neurological and psychiatric disorders and enhancing cognitive function. However, the underlying mechanisms of tDCS are not fully understood, and there is no ground truth for determining, non-invasively, where in the brain tDCS electrical currents flow. At the same time, effective neuronal engagement from tDCS requires accurate localization of the induced currents to specific brain target regions. Machine Learning (ML) has the potential to significantly improve the accuracy and precision of tDCS current prediction. This study aims to develop a novel approach using ML to predict the location and strength of tDCS-induced currents in individual subjects.
Methods: A ring of current-carrying wire wrapped about the circumference of a cylinder containing fluid of high MRI signal is used to establish a theoretical ground truth for ML training, and a spherical phantom with electrodes at opposite ends of its equator is used to evaluate ML models. The cylindrical phantom provides ground truth for both the electric current and the magnetic field induced in the cylinder from the current since the magnetic field is well-known from the Biot-Savart Law, and the electric current is localized to the geometry of the circle of wire with a magnitude that is precisely measurable and controllable. The spherical phantom mimics an ideal scenario closer to the human brain as the current flows through the sphere (volume conduction) rather than being tightly constrained to the wire as in the cylinder case. Different regression models and deep learning techniques, including optimization, are utilized to develop optimal ML models using only data from the cylindrical phantom. ML performance in predicting the magnetic field induced by electric current in the cylinder, sphere, and human subjects is compared to traditional MRI measurements and computational modeling.
Results: The Multi-layer Perceptron (MLP) model emerged as the top-performing regression model in this study, achieving an impressive correlation score of 0.97 between prediction and ground truth on cylinder phantom data after optimization. The optimal MLP model consistently demonstrated strong correlations, reaching a maximum of 0.98, an average of 0.96, and a standard deviation of 0.026 for prediction and computer simulation within the regions of interest (ROI) of spherical phantom data. Additionally, the mean correlation between prediction and MRI measurement was 0.99, indicating they are close to identical. When applied to the human brain, the highest correlation between prediction and computer simulation reached 0.87 in the ROI, with an average correlation of 0.59 and a standard deviation of 0.242. Furthermore, the correlation between our prediction and MRI measurement yielded a maximum value of 0.86, with an average correlation of 0.74 and a standard deviation of 0.166. These findings showcase the excellent performance of our prediction model on ROI from spherical phantom and human brain datasets.
Conclusion: The study demonstrates that a Multi-layer Perceptron Neural Network model trained solely on the cylindrical phantom dataset can effectively predict the tDCS-induced magnetic field in the human brain, reducing noise while preserving the characteristics of MRI measurement. However, the outcomes are influenced by the selection of the brain region of interest, indicating the need for further enhancements to create a comprehensive model that applies to the entire brain. Future work to calculate current density from the magnetic field using Maxwell’s equations will benefit from these enhancements. Nonetheless, this research establishes that machine learning offers a proof of concept of a promising approach for predicting tDCS-induced currents.
Recommended Citation
Olsen, Chikako, "Machine Learning Approach to Predict tDCS-induced Electric Current in the Human Brain" (2023). CUNY Academic Works.
https://academicworks.cuny.edu/cc_etds_theses/1144