Sitemap
A list of all the posts and pages found on the site. For you robots out there, there is an XML version available for digesting as well.
Pages
Posts
publications
Musical signal segmentation method based on self-similarity matrices
Published in 67th Conference on Electrical, Electronic, and Computing Engineering ETRAN, 2023
This paper deals with the problem of automated segmentation and structure analysis of musical pieces on the basis of extracted features from audio recordings of the piece. The complexity of musical piece segmentation problems is a consequence of the multiple elements by which the segments differ, and those can be melody, harmony, rhythm, and timbre. Self-similarity matrices (SSM), based on selected features are used as a cornerstone for segmentation. The features used are MFCC (Mel-Frequency Cepstral Coefficients) and chroma features which follow changes in timbre and harmonic structure. A self-similarity matrix segmentation procedure based on straightforward methods for morphological image processing is proposed. The Dynamic Time Warping algorithm (DTW) was used as a tool for evaluating the degree of similarity of the segments. The proposed procedure was applied to 2 pieces of classical music.
M. Ratković, M. Marijan, T. Miljković, M. Bjelić, "Musical signal segmentation method based on self-similarity matrices" 67th Conference on Electrical, Electronic, and Computing Engineering ETRAN.
A rule-based algorithm for automated renal calculus extraction on CT scans
Published in 67th Conference on Electrical, Electronic, and Computing Engineering ETRAN, 2023
Planning the optimal approach for performing percutaneous nephrolithotomy (PCNL), minimally invasive surgery to remove kidney stones, is critical to reducing the risk of surgical complications. This paper presents the 3D Gastro CT Auto tool for automatic three-dimensional (3D) segmentation and visualization of renal structures with calculus using contrast-enhanced computerized tomography (CT) scans. A novel algorithm for 3D segmentation based on histogram analysis of preprocessed CT scans was implemented as an upgrade of a previously developed 3D Gastro CT Extended application with the aim of simplifying its use, reducing study processing times, and bringing the use of a minimally invasive PCNL planning tool closer to clinical practice. The results of using the tool are shown on the CT scan database of eight patients.
M. Marijan, K. Milićević, O. Durutović, A. Filipović, M. Janković, "A rule-based algorithm for automated renal calculus extraction on CT scans" 67th Conference on Electrical, Electronic, and Computing Engineering ETRAN.
Impact of Protein Representations on Drug–Target Affinity Prediction
Published in 5th Belgrade Bioinformatics Conference, 2024
Accurate and rapid prediction of the binding affinity between potential drug candidates and target proteins can significantly hasten the drug discovery and development process. Utilizing artificial intelligence (AI) models to predict drug-target affinity (DTA) is an affordable and efficient strategy for sifting out undesirable molecules and identifying promising drug candidates. This approach allows researchers to focus on the most promising compounds for further in silico and wet lab experiments, thereby streamlining the overall workflow. Advancements in AI research, such as the development and implementation of graph neural networks (GNN) and attention mechanisms, have significantly improved methods for processing small molecules as potential drug candidates. These developments now allow for very efficient and accurate DTA prediction, without the need for extensive protein processing resources. While this progress marks a significant step forward in computational drug discovery, models that heavily rely on efficient molecule processing may still lack the incorporation of highly specific protein information into their algorithms, which could be crucial for further improvement. In this study, we present a comprehensive analysis of the impact of different protein representations on the accuracy of DTA prediction using two datasets, by implementing and modifying AI models that are based on GNNs and large language models (LLM). Motivated by the intuitive resemblance between traditional motif search methods for protein sequence analysis and conventional one-dimensional convolution in AI signal processing, we propose a protein representation model based on transposed convolutional neural network (NN) layers. Preliminary results indicate that such embeddings improve the overall affinity prediction accuracy, compared to similar models from the literature. Additionally, implementing LLMs to generate protein embeddings independently of other NN layers has demonstrated potential to significantly enhance the accuracy of predicting drug-target pairs that have a very low or unmeasurable affinity.
M. Marijan, I. Tanasijević, "Impact of Protein Representations on Drug–Target Affinity Prediction" 5th Belgrade Bioinformatics Conference.
Message Passing Neural Networks for Sound Source Localization
Published in 33rd Telecommunications Forum TELFOR, 2025
Sound source localization (SSL) is a fundamental problem in audio signal processing. Traditional SSL approaches often degrade in challenging acoustic conditions, while more recent deep learning-based methods struggle to generalize to unseen microphone arrays or to fully exploit the relationships between microphones in an array. We propose Graph-RelNet, a graph-based neural network that extends relation networks with residual graph convolutional layers to regress the direc- tion of arrival (DoA) from generalized cross-correlation with phase transform (GCC-PHAT) features. Experiments on the TIMIT speech corpus demonstrate that Graph-RelNet consis- tently outperforms a regression-based adaptation of GNN-SSL, while also surpassing conventional steered response power (SRP) with delay-and-sum (D&S) beamforming at medium and high signal-to-noise ratios (SNRs), especially in scenarios with few microphones. At 5 microphones and SNR = 30 dB, Graph-RelNet reduces mean absolute DoA error by up to 45%. The model also shows strong generalization to unseen microphone arrays and remains competitive under low-SNR conditions.
M. Marijan, M. Bjelić, "Message Passing Neural Networks for Sound Source Localization" 33rd Telecommunications Forum TELFOR.