Short Bio

Haleh received the B.Sc. and M.Sc. degrees in electrical and biomedical engineering from the Ferdowsi University of Mashhad, Mashhad, Iran, in 2014 and 2017, respectively. She joined the University of Southern California Ph.D. program in Biomedical Engineering as a Viterbi fellow. Her research spans the theory and practice of biomedical signals and imaging processing, machine learning, and data science. Her current research focus is on developing machine learning methods that address the difficulty of applying conventional deep learning models on real-world datasets. More precisely, As a result of her PhD work, She expect to achieve improvement in three major ongoing research areas for unsupervised methods: 1) robustness, 2) generalizability, 3) uncertainty estimation. The newly developed deep learning methods will be used to map anatomical and functional brain changes caused by TBI and from these changes identify distinguishing biomarkers that indicate increased likelihood of the onset of PTE.

Education

University of Southern California

Doctor of Philosophy in Biomedical Engineering • Current

Ferdowsi University of Mashhad

Master of Science in Biomedical Engineering• 2017

Ferdowsi University of Mashhad

Bachelor of Science in Electrical Engineering• 2014

Publications

Robust Variational Autoencoder for Tabular Data with Beta Divergence Haleh Akrami, Sergul Aydore, Richrd M Leahy, Anand A Joshi. ICML UDL2020. [Link]

Brain Lesion Detection Using A Robust Variational Autoencoder and Transfer Learning Haleh Akrami, Anand A Joshi, Jian Li, Sergul Aydore, Richard M Leahy. ISBI 2020. [Link]

Neuroanatomic markers of post-traumatic epilepsy based on magnetic resonance imaging and machine learning Haleh Akrami, Richard M Leahy, Andrei Irimia, Paul E Kim, Christianne Heck, Anand Joshi. medRxiv. [Link]

A Matched Filter Decomposition of fMRI into Resting and Task Components Anand A Joshi, Haleh Akrami, Jian Li, Richard M Leahy. MICCAI 2019. [Link]

Robust Variational Autoencoder Haleh Akrami, Anand A. Joshi, Jian Li, Sergul Aydore, Richard M. Leahy. Arxiv 2019. [Link]

pSConv: A Pre-defined S parse Kernel Based Convolution for Deep CNNs Souvik Kundu, Saurav Prakash, Haleh Akrami, Peter A Beerel, Keith M Chugg. Allerton 2019. [Link]

Lost in music: neural signature of pleasure and its role in modulating attentional resources Samaneh Nemati, Haleh Akrami, Sina Salehi, Hossein Esteky, Sahar Moghimi. Brain Research (2019). [Link]

Group-wise alignment of resting fMRI in space and time Haleh Akrami, Anand A Joshi, Jian Li, Richard M Leahy. SPIE 2019. [Link]

Predicting cognitive scores from resting fMRI data and geometric features of the brain Anand A Joshi, Jian Li, Haleh Akrami, Richard M Leahy. SPIE 2019. [Link]

Culture modulates the brain response to harmonic violations: an eeg study on hierarchical syntactic structure in music Haleh Akrami, Sahar Moghimi. Frontiers in human neuroscience (2018). [Link]

rfDemons: Resting fMRI-Based Cortical Surface Registration Using the BrainSync Transform Anand A Joshi, Jian Li, Minqi Chong, Haleh Akrami, Richard M Leahy. MICCAI 2018. [Link]

Selected Projects

Robust machine learning methods

Python• 2019 - Present

Many machine learning models, especially those based on maximization of log-likelihood,can be easily impacted by outliers in the training data. In this project I develop robust deep learining models using robust statisitc so they can be applied in medicalimaging research with complex training data. Including robust variational autoencoders, robust classifier, robust GAN to an outlier in the dataset[Paper].

Uncertainty Estimation of Autoencoders

Python • 2020 - Present

Determining the uncertainty of the estimates provided by neural networks is critical for clinical applications. We plan to estimate this uncertainty using quantile resgression in the context of the VAE, a popular framework in unsupervised learning.

Transfer learning for improving generalizability

Python • 2019 - 2020

ML methods are based on the assumption that the training and test datasets are sampledfrom the same distribution. However, this assumption may not hold in real-world settings. Leveraging the robustness of the RVAE, I deployed a transfer-learning approach on a test dataset with different characteristics to detect outliers. Our results on MRI datasets demonstrate that we can indeed improve the accuracy of lesion detection by adapting robust statistical models and transfer learning for a variational autoencoder mode [Paper].

Group BrainSync

Python/MATLAB• 2017-2018

I developed a novel method that takes fMRI-based measures of spontaneous brain activity from multiple subjects and applies an orthogonal transform so that they are all temporally ‘aligned’ to an automatically generated averaged template. After alignment, all subjects exhibit approximately equal dynamic brain activity (as would be seen, for example, if comparing multiple subject’s response to a time-locked stimulus)[Paper].

Pre-defined Sparse Kernel Based Convolution for Deep CNNs

Python• 2019

The high demand for computational and storage resources severely impedes the deployment of deep convolutional neural networks (CNNs) in limited resource devices. We Developed a method to reduce CNN model complexity which promises significant improvements in the trade-off between complexity and accuracy for both CNN training and inference. [Paper].

EMG prediction from M1 recordings using group lasso

MATLAB• 2017-Present

To generate a bio mimic movement in paralyzed limbs in spinal cord injured patients developing a controller that replaces spinal cord computation and predict the corresponding muscle activities that generate coordinated movement is necessary. For this purpose, we diploid Generalized Volterra kernel model (GVM) an extension of generalized linear model (GLM) to predict EMG signals based on M1 cortical spike trains during a prehension task.