Paulius Bundza

Mr. Paulius Bundza

orcid

Research Focus & Metrics

Representative Publications

View All Publications
15 February 2026 in Diagnostics
Verified
Journal Article
Open access
PDF file
PBPaulius Bundza
Paulius Bundza
JTJustas Trinkunas
Justas Trinkunas
Background/Objectives: Chest radiography remains a fundamental diagnostic tool for evaluating thoracic disease, yet its interpretation requires considerable time and specialized expertise. Worldwide shortages of trained radiologists can lead to lengthy turnaround times and delayed treatment. This study introduces the Multi-label Chest Abnormality Detection System (MCADS), a deep-learning-driven platform designed to automatically identify and interpret 18 distinct radiographic abnormalities to address these diagnostic challenges. Methods: MCADS integrates a pre-trained DenseNet121 convolutional neural network (via TorchXRayVision) to balance broad pathology coverage with rapid inference. Images are processed asynchronously on a central server to avoid the interruption of clinical workflows. To enhance transparency and clinician confidence, the system employs Gradient-weighted Class Activation Mapping (Grad-CAM) to overlay heatmaps pinpointing image regions most influential to each predicted abnormality. The system was evaluated using eight large, publicly available datasets. Results: When evaluated on diverse datasets, MCADS achieved high area-under-the-curve performance metrics across all 18 target conditions. The platform consistently produced accurate, multi-condition analyses in under thirty seconds per image, demonstrating both reliability and speed suitable for clinical environments. Conclusions: MCADS demonstrates the potential to accelerate chest X-ray interpretation by delivering fast, reliable, and explainable multi-abnormality screening. Its deployment could reduce radiologist workload and mitigate diagnostic delays, offering a pathway to improve patient care within data-driven healthcare environments.

7 April 2025 in Mokslas - Lietuvos ateitis
Verified
Journal Article
Open access
PDF file
PBPaulius Bundza
Paulius Bundza
JTJustas Trinkunas
Justas Trinkunas
Pneumonia detection from chest X-rays is crucial for early diagnosis, and deep learning models –specifically convolutional neural networks (CNNs) – have shown promise in automating this process. In this study, a CNN using the DenseNet-121 architecture was developed and trained, referred to as LDCS2, to classify chest X-ray images as pneumonia or normal, using a combined dataset from three publicly available sources. The CNN approach was chosen over Vision Transformers (ViT) due to lower computational requirements and better performance with limited data. A traditional training, validation, and testing split was used instead of k-fold cross-validation to reduce execution time. LDCS2 demonstrated excellent discrimination between pneumonia and normal images alongside high computational efficiency. These findings highlight the potential of DenseNet-based CNNs for automated pneumonia diagnosis, particularly in resource-constrained settings.