Loading Events

« All Events

  • This event has passed.

Md Navid Akbar’s PhD Dissertation Defense

December 6, 2022 @ 4:00 pm - 5:30 pm

“Inference from Brain Imaging: Incorporating Domain Knowledge and Latent Space Modeling”


Brain imaging can probe the anatomy (structural) of our brain, or its function (functional). A particular imaging modality (unimodal) generally provides only a particular insight into human health. Transcranial magnetic stimulation (TMS), though still in its infancy as a brain imaging modality, is such a functional, unimodal technique. TMS helps model human motor-cortical mapping, using corresponding muscle activity captured by surface electromyography (EMG), but it necessitates a reliable data-driven model. Earlier works have modeled the causal direction only (from cortical representation to muscles), or the inverse direction (from muscles to cortical representation), with simple statistical regression. We modeled this motor-cortical mapping bi-directionally in this dissertation, using deep learning. We first modeled TMS-induced 3D electric field (E-field) in a brain to causal multi-muscle activation picked up by EMG, in a regression task using a convolutional neural network (CNN) autoencoder. By fusing neuroscience domain knowledge (e.g., an empirical neural response profile), we reduced 14% squared error, compared to the baseline model that did not contain this. We then designed our novel inverse imaging CNN model, to reconstruct physiologically meaningful E-field distributions (in the image domain) from a given set of muscle activations (in the sensor domain). By adopting variational inference in the CNN model, to learn the underlying latent space better, we were able to reduce 13% in squared error over our purely CNN baseline.

Diagnosis with brain imaging is often incomplete with a unimodal technique, and having multiple sources (multimodal) may be advantageous. Successful multimodal fusion can provide more holistic information, compared to its constituents. One relevant example is the classification of late post-traumatic seizure (LPTS). Previous works in this space have tackled LPTS classification with either unimodal functional imaging, or non-machine learning (ML) structural modeling. In this dissertation, we first undertook the ML classification of binary LPTS: with unimodal, structural brain imaging, namely diffusion magnetic resonance imaging (dMRI). By incorporating interpretable domain knowledge (post-traumatic lesion volume compensation), we improved 7% in the mean area under the curve (AUC) over the standard technique in literature. Finally, we classified LPTS for a larger sample of subjects, utilizing multimodal imaging, including functional MRI (fMRI) and electroencephalography (EEG). Following unsupervised imputation for any missing modality within the subjects, we introduced our novel multimodal fusion algorithm, which attempts to leverage the underlying structure of the multivariate information. We found that our proposed algorithm improved by 7% in AUC performance, over a naive Bayesian estimator that can handle missing data intrinsically.

Collectively, the work presented here demonstrated that incorporating domain knowledge in the modeling pipeline successfully improved inference. Similar improvements were also observed by learning and leveraging the possible underlying latent structure of the given information, and adapting the models accordingly.


Prof. Deniz Erdogmus (Advisor)

Prof. Mathew Yarossi (Co-advisor)
Prof. Dominique Duncan

Prof. Sarah Ostadabbas


December 6, 2022
4:00 pm - 5:30 pm


Electrical and Computer Engineering
MS/PhD Thesis Defense
PhD, Faculty, Staff