Advancing Electrical Stimulation: Full-Head MRI Segmentation for Abnormal Brain Anatomy with tDCS

Author

Andrew Birnbaum, CUNY City CollegeFollow

Date of Award

2025

Document Type

Thesis

Department

Biomedical Engineering

First Advisor

Lucas C Parra

Keywords

Brain Stimulation, Segmentation, TDCS, Machine Learning, Stroke

Abstract

Evaluating the effectiveness of transcranial direct current stimulation (tDCS) is essential for guiding its integration into therapeutic and performance-enhancement applications. In our laboratory, we investigate the efficacy of tDCS across multiple experimental models, including both animal and human studies. I have contributed significantly to the execution and analysis of these experiments, which include studies in rats and healthy human participants aimed at evaluating whether electrical stimulation of the motor cortex can enhance motor learning. These studies assess improvements in fine motor performance resulting from tDCS. In stroke patients, I contribute to our investigation of tDCS as a rehabilitative intervention, particularly for individuals with apraxia, by running current flow simulations to guide stimulation targeting. This work focuses on brain regions involved in speech production to help restore motor function in the oral muscles.

To optimize stimulation parameters and electrode placement, our laboratory developed ROAST (Realistic vOlumetric-Approach-based Simulator for Transcranial electric stimulation), a software tool for simulating current flow in the brain. However, ROAST's performance is highly dependent on accurate MRI segmentation and head alignment—factors that are often compromised in stroke patients due to abnormal brain anatomy. Existing pipelines like SPM12 and Multipriors frequently fail in such cases, often leading to inaccurate current flow predictions due to misalignment and segmentation errors.

To address these limitations, I led the integration and validation of Multiaxial, a deep learning-based segmentation model originally developed by Lukas Hirsch. Multiaxial employs a consensus approach combining three 2D convolutional neural networks (CNNs) along the sagittal, axial, and coronal planes, followed by a 3D CNN to produce a unified segmentation. Unlike conventional methods, it does not rely on a tissue probability map (TPM), making it robust to anatomical abnormalities seen in clinical populations.

Beyond segmentation, I also focused on solving persistent alignment issues in the ROAST pipeline. I incorporated NiftyReg-based registration to replace the less reliable SPM12 alignment and developed a graphical user interface (GUI) to allow for manual correction of alignment errors. This tool enables users to verify and adjust anatomical landmarks used for cap placement before simulation, ensuring more accurate electrode positioning and reliable current flow predictions, especially in patients with complex brain anatomies.

This dissertation presents my extensive research on using tDCS to enhance motor performance by targeting the motor cortex, with applications spanning motor skill development and post-stroke rehabilitation. It highlights my active involvement in experiments with animal models, healthy individuals, and stroke patients. Central to this work is my contribution to improving current flow simulations by enhancing the ROAST pipeline through the training, validation, and integration of the Multiaxial segmentation model, along with NiftyReg-based and manual alignment tools. I also released an open-source dataset and demonstration code to support future research and clinical applications. These contributions advance the field of neuromodulation by enabling more precise stimulation strategies and paving the way for more effective rehabilitation protocols for stroke patients.

Recommended Citation

Birnbaum, Andrew, "Advancing Electrical Stimulation: Full-Head MRI Segmentation for Abnormal Brain Anatomy with tDCS" (2025). CUNY Academic Works.
https://academicworks.cuny.edu/cc_etds_theses/1248