A team of researchers from the Center for Computer Science of the Faculty of Science and Mathematics, through the EuroCC2/EuroCC4SEE project, with the support of the NCC team of Montenegro, has gained access to the resources of the Leonardo HPC supercomputer. These resources will be used for the efficient development of a system for automatic segmentation of 3D views of mechanical assemblies obtained using 3D scanners. Identification of assembled parts and their relative positions is an important step for reverse engineering, automation of the disassembly process, quality control, AR and VR, etc. These activities are carried out within the project “AI segmentation and inspection by 3D scanning”, in cooperation with partners from France.
Successful application for HPC resources at Leonardo (Benchmark call)
Access to HPC resources is provided for a period of three months, through a successful application to the Benchmark call. During this period, the goal is to demonstrate the ability to efficiently use advanced computing resources, thus earning the right to apply to the Regular call for HPC resources
A new scientific publication by researchers from the University of Donja Gorica and DunavNET explores the innovative use of generative AI in digital agriculture. Titled “Evaluating the FLUX.1 Synthetic Data on YOLOv9 for AI-Powered Poultry Farming”, the study demonstrates how synthetic data, generated using FLUX.1, can effectively enhance deep learning models for chicken detection in smart farms. The paper was published in the Journal of Applied Sciences, a special issue dedicated to the application of computer vision in industry and agriculture [link].
Using generative AI to create sytnhetic data used to train computer vision models for agriculture sector
By combining real and AI-generated images and streamlining annotation with Grounding DINO and SAM2 models, the team achieved impressive detection accuracy—proving that generative AI can bridge the data gap in precision farming. This research is a part of broader effortsincluding PhD research of mr. Stevan Cakic, as well as collaboration with company that produces smart agriculture platform. This was also supported through EuroCC Montenegro initiatives, showcasing how high-performance computing and AI can drive sustainable innovation in agriculture.
High-level architectureused for experiment execution
ABSTRACT – This research explores the role of synthetic data in enhancing the accuracy of deep learning models for automated poultry farm management. A hybrid dataset was created by combining real images of chickens with 400 FLUX.1 [dev] generated synthetic images, aiming to reduce reliance on extensive manual data collection. The YOLOv9 model was trained on various dataset compositions to assess the impact of synthetic data on detection performance. Additionally, automated annotation techniques utilizing Grounding DINO and SAM2 streamlined dataset labeling, significantly reducing manual effort. Experimental results demonstrate that models trained on a balanced combination of real and synthetic images performed comparably to those trained on larger, augmented datasets, confirming the effectiveness of synthetic data in improving model generalization. The best-performing model trained on 300 real and 100 synthetic images achieved mAP = 0.829, while models trained on 100 real and 300 synthetic images reached mAP = 0.820, highlighting the potential of generative AI to bridge data scarcity gaps in precision poultry farming. This study demonstrates that synthetic data can enhance AI-driven poultry monitoring and reduce the importance of collecting real data.
2 Days Hands-On Workshop: Hybrid Learning Experience jointly organized by BioExcel and supported by Sofia University “St. Kliment Ohridski”, Faculty of Chemistry and Pharmacy & Faculty of Physics, DISCOVERER Supercomputer and National Competence Centres in Bulgaria, North Macedonia, Romania, Serbia and Montenegro, this hydrid workshop will offer participants the chance to engage both on-site and online. The workshop will focus on the use of BioExcel core codes such as GROMACS, HADDOCK and PMX with a strong emphasis on hands-on practical sessions and guidance from leading experts in the field.
When & Where: May 21–22, 2025 | Sofia University, Bulgaria & Online
Researchers from the University of Donja Gorica presented a deep learning model for automated tuberculosis detection from chest X-rays at IEEE IT2025 conference. Using a convolutional neural network (CNN), the model classifies images as normal or tuberculosis-positive with an impressive 97.55% accuracy. This breakthrough has the potential to speed up diagnoses, reduce radiologist workload, and improve early detection rates, particularly in low-resource healthcare settings. By leveraging AI for fast and reliable medical imaging analysis, this research highlights the growing role of computer vision in modern healthcare and its ability to enhance efficiency and accuracy in disease detection.
ABSTRACT – This article presents a deep learning model that enables fast and accurate diagnosis of tuberculosis based on chest X-rays. The developed model uses convolutional neural network that enable the automatic classification of chest x-rays into one of two classes: Normal or Tuberculosis with a high degree of accuracy. The model achieved an accuracy of 97.55% on the test data set, indicating its potential to open new perspectives for medical professionals in establishing a tuberculosis diagnosis. This model can significantly speed up the diagnostic process, reducing the workload of medical workers and increasing their productivity in the fight against tuberculosis, one of the most common lung diseases.
At the IT2025 IEEE Conference in Žabljak, researchers from the University of Donja Gorica presented a study on voice cloning and text-to-speech (TTS) technology for cultural heritage preservation. Their research compared state-of-the-art AI models, including Realtime Voice Cloning (RVC), Tortoise AI, Bark, and Coqui AI, to evaluate how small, high-quality datasets can produce more accurate and natural-sounding speech than large, unstructured ones. The study highlights the potential of AI in preserving the Montenegrin language and oral traditions, enabling the creation of audiobooks, digital archives, and interactive experiences. This research paves the way for more accessible educational resources and enhanced cultural engagement using AI-driven speech synthesis.
ABSTRACT – This research presents a comparative analysis of modern voice cloning systems, focusing on their ability to generate high-quality speech from limited training data. The paper aims to demonstrate that carefully curated smaller datasets can produce superior results to larger, less structured datasets. The investigation of multiple state-of-the-art models, including Realtime Voice Cloning (RVC), Tortoise AI, Bark, and Coqui AI, establishes optimal data preparation protocols and identifies critical factors in training data quality, with particular emphasis on applications for the Montenegrin language and cultural preservation.
At the IT2025 IEEE Conference in Žabljak, researchers from the University of Donja Gorica presented their latest study on using Artificial Intelligence (AI) for breast cancer diagnostics. The research explores the application of deep learning models, ResNet152 and DenseNet121, to analyze mammographic images. Beyond the clinical results, the study emphasizes the implications of leveraging high-performance computing (HPC) infrastructure to optimize model training and evaluation. By porting the experimental setup to HPC resources, the research opens pathways for faster development cycles, the exploration of more complex architectures, and scalability for real-world implementation.
ABSTRACT – Artificial Intelligence is rapidly advancing the medical field by providing innovative disease diagnosis, treatment, and research approaches. This study explores the application of artificial intelligence in breast cancer diagnostics, focusing on using convolutional neural networks and deep learning to analyze mammographic images. ResNet152 and DenseNet121 models were used to classify malignant changes, achieving AUC scores exceeding 0.9, demonstrating their clinical utility. The research emphasizes how artificial intelligence can enhance screening efficiency, expedite diagnostic processes, and facilitate personalized treatment approaches. Ethical considerations, including patient safety and the transparency of artificial intelligence systems, were also analyzed. The findings underscore the potential of artificial intelligence to transform diagnostic procedures for breast cancer and highlight the importance of further research to integrate these technologies into clinical practice.
Uhura Solutions is developing AI platform for document-driven process automation in financial services. The “Generative AI Intelligent Process Automation Platform – GAIPAP” project is a transformative initiative aimed at revolutionizing the financial industry through the integration of advanced AI-driven automation solutions that combine capabilities of Large Language Model (LLM), low-code development, and process automation workflows. This unique fusion of technologies holds the potential to significantly enhance operational efficiency, cost reduction, and customer experience within the financial sector.
An exceptional aspect of the platform is the incorporation of fine-tuned Large language models (LLMs), which sets it apart from generic AI solutions. These LLMs, refined for the financial industry’s unique language, context, and patterns, promise to deliver more precise and relevant insights. The platform’s value proposition is further strengthened by its ability to streamline workflows, offer scalability, real-time data insights, and elevate customer experiences. The integration of a low-code development framework and process automation workflows enables swift prototyping and deployment of AI-driven models, thereby accelerating time-to-market and capitalizing on market opportunities effectively.
The project involves the use of open-source Python libraries and pre-trained Large language models which are fine-tuned on a custom-made private dataset created by the team, allowing for specialized tasks within the financial sector. Leonardo supercomputer enabled us to make a leap with GPU training experiments from 1B to 7B and higher Large language models. We are focusing on various LLM research and optimization methods in this phase of the project, which include hyperparameter tuning, model quantization and pruning, and parameter efficient fine-tuning (PEFT). Special attention is put on dataset preprocessing including quality filtering, deduplication and privacy redacting.
This project, developed and managed by Uhura Solutions, was awarded with 4,500 node hours of GPU resources (8x64GB) on the Leonardo Buster HPC for a duration of 12 months.
