About
I am a second year PhD student at FIIT STU in Bratislava, under the supervision of Prof. Wanda Benešová. I view the processing of heterogeneous sources of information, modalities, as a crucial step for advancing the universality of deep learning solutions. While originally this concept aimed to simulate the 5 human senses, I focus on applying it to combining various visual sensors for complex scene understanding. I believe that with the recent advancements in attention-based neural network architectures, we can develop high-performance explainable solutions for even the most complex tasks.
My research focuses on multimodal deep learning applications of computer vision. I apply attention mechanisms to rectify and fuse visual modalities to advance the performance of neural networks in real-world perception tasks. I’m currently working on the advanced application of these systems which allow fast and accurate object detection and trajectory prediction in autonomous vehicles.
Research
For a complete and up-to-date list of all my publications go to my Google Scholar page
-
Halinkovic, Matej, et al. “Intrinsically explainable deep learning architecture for semantic segmentation of histological structures in heart tissue.” Computers in Biology and Medicine 177 (2024): 108624.
Abstract: Background: Analysis of structures contained in tissue samples and the relevant contextual information is of utmost importance to histopathologists during diagnosis. Cardiac biopsies require in-depth analysis of the relationships between biological structures. Statistical measures are insufficient for determining a model’s viability and applicability in the diagnostic process. A deeper understanding of predictions is necessary in order to support histopathologists. Methods: We propose a method for providing supporting information in the form of segmentation of histological structures to histopathologists based on these principles. The proposed method utilizes nuclei type and density information in addition to standard image input provided at two different zoom levels for the semantic segmentation of blood vessels, inflammation, and endocardium in heart tissue. Results: The proposed method was able to reach state-of-the-art segmentation results. The overall quality and viability of the predictions was qualitatively evaluated by two pathologists and a histotechnologist. Conclusions: The decision process of the proposed deep learning model utilizes the provided information sources correctly and simulates the decision process of histopathologists via the usage of a custom-designed attention gate that provides a combination of spatial and encoder attention mechanisms. The implementation is available at https://github.com/mathali/IEDL-segmentation-of-heart-tissue
-
Sloboda, T., Hudec, L., Halinkovic, M. and Benesova, W., 2024. Attention-Enhanced Unpaired xAI-GANs for Transformation of Histological Stain Images. Journal of Imaging, 10(2), p.32.
Abstract: In the field of heart transplantation, the ability to accurately and promptly diagnose cardiac allograft rejection Histological staining is the primary method for confirming cancer diagnoses, but certain types, such as p63 staining, can be expensive and potentially damaging to tissues. In our research, we innovate by generating p63-stained images from H&E-stained slides for metaplastic breast cancer. This is a crucial development, considering the high costs and tissue risks associated with direct p63 staining. Our approach employs an advanced CycleGAN architecture, xAI-CycleGAN, enhanced with context-based loss to maintain structural integrity. The inclusion of convolutional attention in our model distinguishes between structural and color details more effectively, thus significantly enhancing the visual quality of the results. This approach shows a marked improvement over the base xAI-CycleGAN and standard CycleGAN models, offering the benefits of a more compact network and faster training even with the inclusion of attention.
-
Kveton, M., Hudec, L., Vykopal, I., Halinkovic, M., Laco, M., Felsoova, A., Benesova, W. and Fabian, O., 2023. Digital Pathology in Cardiac Transplant Diagnostics: From Biopsies to Algorithms. Cardiovascular Pathology, p.107587.
Abstract: In the field of heart transplantation, the ability to accurately and promptly diagnose cardiac allograft rejection is crucial. This comprehensive review explores the transformative role of digital pathology and computational pathology, especially through machine learning, in this critical domain. These methodologies harness large datasets to extract subtle patterns and valuable information that extend beyond human perceptual capabilities, potentially enhancing diagnostic outcomes. Current research indicates that these computer-based systems could offer accuracy and performance matching, or even exceeding, that of expert pathologists, thereby introducing more objectivity and reducing observer variability. Despite promising results, several challenges such as limited sample sizes, diverse data sources, and the absence of standardized protocols pose significant barriers to the widespread adoption of these techniques. The future of digital pathology in heart transplantation diagnostics depends on utilizing larger, more diverse patient cohorts, standardizing data collection, processing, and evaluation protocols, and fostering collaborative research efforts. The integration of various data types, including clinical, demographic, and imaging information, could further refine diagnostic precision. As researchers address these challenges and promote collaborative efforts, digital pathology has the potential to become an integral part of clinical practice, ultimately improving patient care in heart transplantation.
-
Halinkovic, M., Mušková, K., Sloboda, T., Lepá?ek, M., Kan?árová, H., Ries, M. and Prnová, M.Š., 2024. MLtox, online phototoxicity prediction webpage. Toxicology in Vitro, 94, p.105701.
Abstract: Phototoxicity, sometimes in the literature referred to as photo-irritation, is a chemically induced reaction requiring light. While it is generally accepted that phototoxicity testing can be performed in the majority of cases in vitro (i.e. without the use of experimental animals), these tests may sometimes provide contradictory predictions. Understanding the mechanisms of initiating events based on the molecule's structure and its ability to reach the excited state and consequently generate ROS enables the creation of predictive QSAR for this adverse outcome. The ability to predict the phototoxicity potential via a QSAR model is beneficial in reducing the number of mechanical in vitro/in chemico tests needed to demonstrate absence of phototoxicity and it is very helpful in the overall safety assessment process. The QSAR prediction model presented here focused on developing a robust platform freely available on the web via the link http://mltox.online to provide interpretable predictions of the phototoxicity of tested molecules. Great attention was devoted to interpretability and explainability of the prediction results. The web application allows the user to input a chemical by CAS number, SMILES code or trivial name. The user can choose between simple prediction or advanced tools options. These extended tools include the artificial intelligence explainability of model prediction using XSMILES (interactive visualization technique to support the interpretation of SMILES) and SHAP values (impact each element on the prediction). The comprehensive tools in question allow the user to explore the properties of phototoxic substances and to understand the prediction outcomes better.
-
Halinkovic, M. and Benesova, W., 2022, August. SpringNet: A Novel Deep Neural Network Architecture for Histopathological Image Analysis. In The International Conference on Innovations in Computing Research (pp. 65-75). Cham: Springer International Publishing.
Abstract: Reliable and capable tools are needed for segmentation and classification of nuclei in histopathology images to improve the workflow of histopathologists and lessen the burden of diagnosis. Analyzing histopathology images can be challenging for deep learning models due to the various sources of noise. We propose a pipeline for segmenting and classifying nuclei, trained on a combination of datasets to ensure robustness. This makes our pipeline capable of generalizing on image tiles with varying Hematoxylin & Eosin concentrations. For nuclei classification, we present a novel convolutional neural network classifier that demonstrate state-of-the-art classification performance on the Lizard dataset.
Theses
- Research methods of using deep neural networks in computer vision
Halinkovic, Matej. Dissertation thesis: FIIT STU, ongoing - Interpretability and explainability of deep learning systems in the domain of microscopic medical imaging
Halinkovic, Matej. Diploma thesis. Bratislava: FIIT STU, 2023. - Music Analysis Using Artificial Neural Networks
Halinkovic, Matej. Bachelor thesis. Bratislava: FIIT STU, 2021.
Teaching
- Digital Processing of Sound, Image, and Biosignals – Teaching Assistant
- Neural Networks – Teaching Assistant