1 Investigating PET Image Enhancement through Neural and Classical Optimal Transport Methods
Emir Türkölmez
Master Student
(Supervisor: Prof.Dr.Selim Aksoy) Computer Engineering Department
Bilkent University
Abstract: Positron Emission Tomography (PET) provides critical functional insights for diagnosing conditions like cancer and neurological disorders but is hindered by low spatial resolution, limiting its ability to capture fine structural details. This study explores two methods to enhance PET image quality using their paired and registered high-resolution Magnetic Resonance (MR) image counterparts. Classical transport techniques, such as Earth Mover’s Distance (EMD), solve linear optimization problems to align the PET intensity distribution with MR anatomical structures, offering precise and interpretable results. Neural optimal transport, on the other hand, is based on nonlinearity to learn complex, feature-rich mappings that capture semantic relationships. Through experiments on paired PET-MR datasets, we demonstrate that classical methods provide immediate, reliable transformations, while neural approaches offer the potential to capture intricate feature-level details. These complementary techniques highlight a promising path for improving PET image quality and enhancing its diagnostic utility in clinical practice. Keywords: Positron emission tomography, magnetic resonance imaging, image enhancement, neural networks, optimal transport, Earth Mover’s Distance
DATE: December 09, Monday @ 15:30 Place: EA 502
2 Learning Based Time-Optimal Motion Control
İlyas Kocaer Master Student
(Supervisor: Asst.Prof.Özgür Salih Öğüz) Computer Engineering Department
Bilkent University
Abstract: Motion control systems in distinct industrial systems can bear various responsibilities and objectives where time-optimality is one of the possible objectives and presently unresolved problem. A hybrid motion control system architecture which is formed with stochastic and deterministic components in a cascaded structure for time-optimality purpose will be presented by also discussing robustness against possible variations which might occur in the life-time of a system and between manufactured systems, accuracy requirements to ensure settling position windows without overshoot or undershoot behaviors comparing with the reference command signal and compensation capabilities for existing disturbances. Possible realizations of corresponding physical systems and the system stability requirements with verification techniques will be evaluated.
DATE: December 09, Monday @ 15:50 Place: EA 502
3 Predicting the Next Generation of Key Developers in Software Projects
Fereshteh Vedadi
Ph.D. Student
(Supervisor: Asst.Prof.Eray Tüzün) Computer Engineering Department
Bilkent University
Abstract: In software projects, certain developers hold significant influence either by overseeing the entire project, focusing on specific components, or facilitating communication and coordination among teams. These key developers are vital in ensuring the project’s overall success and ongoing maintenance. Given the significance of these individuals, proactively predicting them enables organizations to gain a competitive edge by nurturing their future talents. Numerous studies have focused on identifying a developer’s current role in software projects across various contexts. However, there are limited investigations into forecasting a developer’s future role within a project. This study addresses this gap by proposing a methodology to predict future key developers of a project based on developers’ initial activities. Our approach leverages four established predictive models: k-Nearest Neighbors, Logistic Regression, Random Forest, and Naïve Bayes, trained on a set of features inspired by sports analytics. These features are analogous to those used to assess athletes’ performance in different kinds of sports and emphasize the evaluation of developers’ technical proficiency and collaborative dynamics within their first six months of contribution to software projects. We assessed our methodology across four open-source projects: Vuejs-Core, Spring-Security, Moby, and Gitea. Based on the obtained results, the proposed methodology, employing Random Forest, proved to be the most effective model, achieving up to 71.18% F1 Score in predicting key developers based on their historical performance and collaboration metrics utilizing two different sources of ground truth data. By enhancing the prediction accuracy, our approach not only provides a framework for organizations to predict and cultivate future key developers proactively but also supports the long-term success and sustainability of software projects.
DATE: December 9, Monday @ 16:10 Place: EA 502