Research

Anti-cancer Drug sensitivity prediction (Ph.D.):

My Ph.D. is mainly focused on developing machine learning and deep learning tools for anti-cancer drug sensitivity predictions which is an intersection of precision medicine and machine learning. In precision medicine, high dimensional omics data (DNA-seq, RNA-seq, Protein-seq) is being used heavily along with different type of drug descriptors. Handling omics data is very challenging as its dimension is regularly in the range of 20k. Therefore using shallow learning models provides mediocre accuracy. In my Ph.D., I focused on developing REFINED (REpresentation of Features as Images with Neighborhood Dependencies) for training Convolutional Neural Networks. REFINED converts high dimensional data, unsupervised, into compact images such that similar features (i.e. genes) are in close neighborhood and dissimilar features are far away within the created image. REFINED produces a unique coordinate in a 2D space for each feature. The published papers from my work are listed here for your viewing pleasure:

• REFINED (REpresentation of Features as Images with Neighborhood Dependencies) Paper

• Integrated REFINED Paper

• Statistical transfer learning in pharmacological studies

Emotion (M.S. & Ph.D.):

Emotion recognition is a very challenging task, as the emotional states are associated with wide variety of human feelings, thoughts and behaviors. In this research, which was a long collaboration lasted from my M.Sc. to Ph.D. , we decomposed EEG signals from the DEAP dataset using wavelet transform into different frequency bands. Extracted features and transformed them using PCA to train an SVM classifier. Using this system we classified the valence and arousal level of the emotional states of the participants. Each specific weighted combination of valence and arousal will represents an emotional state. The complete description of our work is provided in the below paper.

• Emotion recognition with machine learning Paper

Active shooter detection (Ph.D.):

In this study, we designed a cyber-physical framework that analyzes RF micro-Doppler and Range-Doppler signatures for potential active shooter detection. From a machine learning perspective, the RF sensor spectrogram images are conducible for training using deep convolutional neural networks. However, generating a large training dataset with exhaustive variety of multi-person scenarios is extremely time consuming and nearly impossible due to the wide range of combinations possible. We present approaches for multi-person spectrogram generation based on individual person spectrograms that can augment the training dataset and increase the accuracy of prediction. The proposed system can aid as a standalone or complemented by a video surveillance tool for anomaly detection in scenarios involving single or multiple individuals.

• Active shooter detection with Machine Vision Paper

Parkinson tremor assessment (M.S. & Ph.D.):

My M.Sc. research was focused on designing a classification system for automatically assessment of Parkinson's disease patients tremor based on UPDRS metrics. We used a smartphone to acquire acceleration data of PD patients hand tremor at Rasoul Akram hospital of Iran medical sciences university. The designed system was implemented on a PCB board with an ARM7 processor. Different aspects of my work and collaboration with other universities were published in conference proceedings and peer- reviewed journals. Publications are listed:

• Neural network system Paper

• Classification system and hardware implementation Paper

• Tremor and gait classification Paper

Magnetic Resonance Spectroscopy (Ph.D.):

First year of my Ph.D. research was focused on the quantification of brain's metabolite concentration using magnetic resonance spectroscopy (MRS). MRS signal analysis requires multiple pre-processing steps such as separating macro-molecules from metabolites that is called baseline correction. To address this issue I designed an automatic baseline correction algorithm.

There are two types of MRS signal analysis frameworks, black-box method and basis set methods. In black box methods, the metabolite signals are detected in data-driven manner, whereas in basis set methods they are identified using prior knowledge. In my work, I developed a methods that integrates both approaches and quantifies the brain's metabolites concentration.

• Brain metabolite's concentration quantification Paper

• MRS Baseline correction Paper

3D MRI segmentation (Genentech):

• Kidney segmentation using 3D U-Net localized with Expectation Maximization Paper