Informatics Engineering

Topic title	Possible scientific supervisors	Source of funding
Early Sepsis Prediction from Heterogeneous and Imbalanced Clinical Data Using Causal Self-Supervised Learning Methods	prof. dr. Rytis Maskeliūnas	state-funded	Research Topic Summary. The aim of this thesis topic is to develop an advanced early sepsis prediction system based on deep and self-supervised learning principles, capable of processing clinically heterogeneous, irregularly sampled, and incomplete time series data. Since real-world intensive care databases often lack reliable annotations, this dissertation proposes the use of a causal self-supervised learning paradigm that enables the discovery of temporal causal patterns of health state evolution without direct dependence on labeled data. The model architecture may be based on temporal contrastive representation synthesis, causal latent state modeling, and uncertainty-aware reweighting of imbalanced samples, which would not only improve the accuracy of sepsis risk estimation but also provide calibrated temporal uncertainty assessments to dynamically evaluate prediction reliability over time. Additionally, the research should explore multimodal data integration (e.g., physiological signals, laboratory results, and clinical notes) through multimodal contrastive learning, aiming to construct a unified, semantically interpretable patient state representation map. The developed system would not only predict the onset of sepsis but could also serve as a real-time early warning tool, identifying causal risk factors and their temporal dynamics.
Penetration testing based method for assessing the cyber resilience of the Internet of Things (IoT) to quantum threats	prof. dr. Algimantas Venčkauskas	state-funded	Research Topic Summary. The topic is relevant because assessing the cyber resilience of IoT is a complex problem due to the heterogeneity of the IoT ??environment, the large number of different components and protocols, especially due to the threats of quantum computing. The life cycle of IoT systems is long and complex: industrial and medical IoT devices have been in use for decades. The transition to hybrid systems, quantum-safe cryptography (Post-Quantum Cryptography, PQC) is slow and complex. It is necessary to identify the most vulnerable devices and protocols as soon as possible. Penetration testing of IoT systems, using the capabilities of artificial intelligence, is a promising way to solve this problem. Penetration testing of IoT systems (penetration testing, pentest) differs from traditional IT systems, is complex due to the heterogeneity of the IoT ??environment, the large number of different components and protocols. A realistic transition to quantum security requires hybrid cryptosystems (using PQC together with traditional cryptography). The dissertation would investigate how these hybrid systems operate in resource-constrained IoT environments and whether they themselves create new traditional vulnerabilities (e.g., excessive energy consumption or slow connection establishment leading to DoS attacks).
Multimodal segmentation of transparent and reflective objects for real-time industrial systems	prof. dr. Armantas Ostreika	state-funded	Research Topic Summary. This project targets reliable separation of transparent and mirror-like objects in industrial scenes, where conventional vision often fails due to reflections and unreliable depth. We will develop a real-time method that fuses color, depth, and polarization cues, grounded in the physical laws of refraction and reflection. A reference demonstrator will be built to run on local computing devices and integrate with a robotic gripper or conveyor line. Evaluation will cover not only accuracy and false-alarm reduction, but also latency, processing throughput, energy consumption, and robustness across different sensors. The outcome is open-access software and clear deployment guidelines (illumination setup, sensor selection, calibration) that improve quality inspection accuracy, reduce downtime, and enhance operational safety.
Multimodal Deep Learning Architecture for Prognosis and Analysis of Laryngeal Cancer Treatment Outcomes using MRIgRT Data	prof. dr. Rytis Maskeliūnas	state-funded	Research Topic Summary. The doctoral thesis topic is aimed at the development and investigation of an artificial intelligence (AI) system based on computer vision for analyzing the condition and treatment progress of patients with laryngeal cancer. The system will be based on MRI-guided radiotherapy (MRIgRT) data acquired using the Elekta Unity system at the Hospital of Lithuanian University of Health Sciences (LSMU Kauno Klinikos). The study will utilize DICOM MRI datasets and anonymized clinical information, including disease-specific questionnaires and video-endoscopic recordings of tumor tissues, collected before and during treatment. The goal is to quantitatively assess tumor response, early and late treatment effects, and the structural and functional changes in surrounding soft tissues (e.g., acoustic voice parameters). The research aims to develop a predictive AI model integrating radiological, video-endoscopic, acoustic, and questionnaire-based data to forecast individualized treatment outcomes across different therapeutic modalities (for example, such as conventional radiotherapy, MRI–Linac radiotherapy, and transoral laser microsurgery (TOLMS)).
Research of methods for multi-spectral image reconstruction from RGB image	prof. dr. Vidas Raudonis	state-funded	Research Topic Summary. The main aim of the research is to create an "Optical Soft Sensor," an AI model relying on deep learning architectures to accurately reconstruct high-resolution multispectral images from low-cost, ordinary RGB camera data. This project is highly relevant because multispectral imaging provides access to otherwise "invisible" analytical information for fields like precision agriculture, geology, and biomedical diagnostics, ultimately increasing the accessibility of these advanced monitoring techniques.
Using EEG-EMG Approaches and Deep Learning Techniques for Emotion Recognition and Prediction in VR Serious Games	prof. dr. Robertas Damaševičius	state-funded
Development and research of a hybrid optimisation method for robot control using an AI platform	prof. dr. Renaldas Urniežius	state-funded	Research Topic Summary. We are creating a new generation of artificial intelligence by combining data science with fundamental laws of physics. Traditional AI systems learn only from data, which requires enormous resources and sometimes leads to errors or illogical solutions. Our hybrid method has "taught" AI to understand basic rules. This allows us to develop AI platforms that operate significantly faster, require less data, and make better, safer, and more physically realistic decisions. If you are not afraid of difficulties and challenges, join us in our efforts to create more reliable autonomous systems (e.g., https://www.youtube.com/watch?v=DvgsmHiadhQ), smarter robots, and a more efficient industry.
Method for identifying and classifying malicious social media messages	prof. dr. Jevgenijus Toldinas	state-funded	Research Topic Summary. Scientific problem. The main means of communication in the modern world are social networks, which are full of harmful messages that can cause both psychological and financial harm. Most websites lack services that automatically delete or return malicious information to the sender for correction or notify the sender about inaccurate message content. Implementing such systems could involve the use of methods for identifying and classifying malicious messages. The objective is to propose and study a method that identifies and classifies malicious messages on social networks.
Advancing Cyber Resilience for Secure Digital Ecosystems with a risk based assurance methodology	prof. dr. Šarūnas Grigaliūnas	state-funded	Research Topic Summary. The topic strengthens cyber resilience of digital ecosystems using a risk based assurance methodology across the full lifecycle. It blends threat modelling, supply chain vetting, resilience metrics and governed recovery to help organizations withstand incidents and recover faster. The outcome practical guidelines and demonstrators for business, public sector and industry.
Integrating Explainability in Vision-Language Models: Methods and Applications for Multimodal AI Transparency with Application in Furniture Manufacturing and Surface Defect Analysis	doc. dr. Gintaras Dervinis	state-funded	Research Topic Summary. The integration of Artificial Intelligence (AI) and Machine Learning algorithms into classical industrial automated and robotic production systems makes the activities of industries more efficient, providing the opportunity not only to fully automate processes and improve decision-making. In furniture production, the application of advanced AI models has shown very high potential in production quality control, especially in detecting and identifying surface defects. Using vision-language models (VLM) that combine visual and textual data, it is possible to fully or partially automate the quality inspection process and identify various types of defects. However, the “black box” nature of these models poses a challenge: there is a lack of transparency and validity of the results created/generated. To solve this problem, it is necessary to integrate explainability methods that provide insights into the decision-making process and increase trust in AI systems.The application of modern VLM, such as CLIP and BLIP, has shown the ability to jointly understand visual and textual input data. These models use both visual and textual information. Although they recognize patterns/textures with reasonable accuracy and generate consistent descriptions, their decision-making processes are often unclear. The lack of explainability makes it difficult for users to understand why certain predictions or classifications are made, especially when the model identifies/detects a surface defect or flaw. The biggest challenge in assessing the clarity, usefulness, and accuracy of explanations is the development of objective metrics. The integration of explainable intelligence (XAI) into VLM is essential for their application in relevant application areas, such as surface defect analysis, e.g. in furniture manufacturing. The goal is to advance the XAI field by exploring new methods for explaining VLM, developing domain-specific datasets, and interactive tools to improve user understanding.
Personalized non-invasive estimation of glucose dynamics from finger optical signals	prof. dr. Armantas Ostreika	state-funded	Research Topic Summary. The project aims to develop a personalized, non-invasive system for assessing glucose trends from finger optical signals and detecting high-risk episodes, reducing the need for frequent finger pricks. An optical probe with multi-wavelength emitters and a photodiode will be built, with integrated temperature and contact-pressure sensing and a stable mechanical holder. Software will handle signal-quality control, artifact suppression, and uncertainty estimation, while personalized learning will use participants’ reference measurements. Evaluation will cover trend-error, episode detection, latency, throughput, and energy consumption. The outcome is a working prototype, open-access methods, and clear deployment guidelines.
Developing a Post-Quantum Resilience Method for Cyber Threat Management	prof. dr. Šarūnas Grigaliūnas	state-funded	Research Topic Summary. The dissertation focuses on the emerging cybersecurity challenges posed by quantum computing. Its aim is to develop a post-quantum resilience method enabling organisations to migrate safely to PQC technologies, detect deployment-related risks and ensure operational continuity in the quantum-threat landscape. The research integrates post-quantum cryptography (CRYSTALS-Kyber, FALCON), threat intelligence, ML/AI-based detection of PQC implementation errors, and crypto-agility practices applicable to IT, OT, cloud and microservice environments. The resulting method will be experimentally validated and delivered as a practical framework and set of guidelines for secure PQC migration and cyber-resilience management.