Transformers have gained prominence in natural language processing due to their representational capabilities and performances. Transformers process natural language as a sequence on finite context windows; however, global relationships among words beyond these windows cannot be completely modeled via sequence processing only. Graph neural network (GNN) based models have been proposed to alleviate this problem, as they provide geometric extensions to neural networks, enabling models to learn associations within a text. However, regular graph-based methods ignore the sequential nature of underlying texts. In this paper, we propose EmoVis, the first generic graph-based neural network that utilizes visibility graphs, which converts classical time-series information to graph representations. We cast the problem as an emotion classification task, enabling the proposed model to learn associations between the labels and words in a sentence. Moreover, EmoVis can be used as a highly modular graph-based extension to any transformer-based model, significantly improving their performance and learning capabilities in various languages. We experimentally show that EmoVis enables transformer-based models to outperform the state-of-the-art baselines across three diverse datasets in different languages in the SemEval2018 competition datasets and the GoEmotions dataset.

The gSlider, a state-of-the-art diffusion MRI technique, requires extensive sampling of thick-slice RF-encoding and q-space points to resolve fiber structures at the submillimeter level. This results in impractically long scan times. Our goal is to super-resolve gSlider data in RF-encoding and q-space dimensions to enhance scan efficiency while maintaining high-fidelity diffusion metrics. We introduce a novel self-supervised model, sq-QUCCI, that cascades two neural field modules with physics-driven regularization to super-resolve undersampled gSlider acquisitions. Compared to state-of-the-art baselines, sq-QUCCI achieves superior fidelity in diffusion metrics at 500-μm resolution while enabling 7-fold reduction in scan time for gSlider acquisitions. sq-QUCCI enables collection of whole-brain high-spatial/angular-resolution, high-SNR diffusion MRI data in a 15-min scan by super-resolving across RF-encoding and q-space dimensions of undersampled gSlider acquisitions, overcoming the efficiency barrier for adoption in clinical settings.

In magnetic particle imaging (MPI), selection field (SF) gradients are utilized to form a field-free point (FFP) in space, such that only the magnetic nanoparticles (MNPs) in the vicinity of the FFP respond to the applied drive field (DF) and contribute to the received signal. While the relaxation behavior of MNPs adversely affects image quality by reducing signal intensity and causing blurring, it also provides MPI with functional imaging capabilities, such as viscosity and temperature mapping. This work investigates the effects of SF gradients on the relaxation behavior of the MNPs using an in-house magnetic particle spectrometer (MPS) setup equipped with additional DC electromagnet SF coils, which switches the MPS setup into an MPI system. The results reveal that the presence of SF gradients boosts the viscosity sensitivity of MPI, and that the MPI signal can be sensitized to viscosity even at high DF frequencies and amplitudes if sufficiently large SF gradients are applied.

This study examines a deep learning method for EPI susceptibility correction that merges physics-constrained learning with phase-injected formulation and two-step training, and assesses its impact on diffusion metrics. It offers quick and effective correction, suitable for clinical and research use.

Diffusion MRI (dMRI) has enabled the study of the brain’s complex architecture in vivo. However, resolving fiber structures at a finer detail necessitates denser q-space sampling, which can be impractical due to excessively long scan times . Deep learning methods for angular q-space interpolation show potential for reduced scan time by enabling q-space downsampling. However, these methods often demand large training data and pose challenges when dealing with non-conforming data, such as in pathological cases.

We propose a subject-specific, unsupervised Q-space Upsampling via physics-Constrained Coordinate-based Implicit network (QUCCI) that angularly interpolates q-space data. QUCCI combines implicit multilayer perceptron regularization with physics-driven spherical harmonic (SH) regularization and image regularization in the context of dMRI. In comparison to traditional and other deep learning methods, QUCCI demonstrates superior angular interpolation for highly undersampled q-space acquisitions.

In Magnetic Particle Imaging (MPI), the distribution of magnetic nanoparticles (MNPs) is imaged by moving a field free point (FFP) in space. All MNPs in close vicinity of the FFP contribute to the signal induced on the receive coil. The relaxation behavior of these MNPs are subject to a DC field due to the selection field (SF). In this work, we investigate the effects of the DC field on the relaxation behavior of the MNPs, with the goal of understanding the differences between the measured relaxations in Magnetic Particle Spectrometer (MPS) setups vs. MPI scanners.

UC Berkeley EECS Departmental Fellowship
Monthly stipend and accommodation support during M.Sc. by Scientific and Technological Research Council of Turkey (TUBITAK)
Full tuition waiver and stipend during M.Sc. by Bilkent University

Research Excellence Award
Social Justice and Sustainability Award
High Honor Student at Bilkent University
Bilkent University Merit Scholarship
Summa Cum Laude

International Genetically Engineered Machine (iGEM) golden award (international synthetic biology competition organized by iGEM Foundation, Boston, USA).
METU High-school Merit Scholarship