We are pleased to congratulate Aleksas Vaitulevičius, the new Doctor of Natural Sciences in the field of Informatics, who has successfully defended his doctoral dissertation “Research and Modeling of Digital Signals for the Detection of Cancerous Areas of Tissue in Different MRI Modalities”.
Dr. Aleksas Vaitulevičius and Dissertation Defense Board
Aleksas Vaitulevičius’s research may be of interest to researchers in artificial intelligence and data science, medical image analysis specialists, radiologists, and other scientists interested in the application of computational methods for prostate cancer diagnostics.
The dissertation investigates the detection and segmentation of prostate cancer in multiparametric magnetic resonance imaging (mpMRI) using deep learning methods, with a particular focus on nnU-Net. It also examines prostate zone segmentation, as previous research has shown that incorporating prostate zonal information can improve prostate cancer segmentation and detection accuracy.
Key results include:
- Two novel prostate zone segmentation workflows based on deep neural networks trained using nnU-Net were proposed. The most accurate results were achieved by using separate models for the transition zone and peripheral zone segmentation.
- Several U-Net variants employing attention mechanisms were evaluated for the prostate zone segmentation task, but no statistically significant differences between the compared neural network architectures were found.
- Four novel prostate cancer segmentation ensembles using nnU-Net-trained deep neural networks and prostate zone masks were proposed. The most accurate results were achieved by combining a model specialised in segmenting lesions in the peripheral zone with a model, which segments prostate cancer in the whole prostate gland and uses prostate zone masks as a separate input channel.
- A modified Dice Score Coefficient (DSC) metric – the tolerant Dice Score Coefficient (DSCτ) – was proposed to account for high inter-observer variability in the ground-truth segmentations of prostate cancer lesions. The experiments showed that DSCτ consistently yields higher scores than conventional DSC, indicating that it accounts for tolerable discrepancies in segmentation boundaries.
The prostate zone segmentation experiments used two publicly available datasets, ProstateX and the Decathlon Challenge dataset. The prostate cancer segmentation experiments were conducted using a non-public dataset provided by the National Cancer Institute.
The dissertation was conducted during 2021-2025 at the Faculty of Mathematics and Informatics, Vilnius University, which carries out doctoral studies in informatics, informatics engineering, and mathematics science fields.
Scientific supervisor – Prof. Dr. Povilas Treigys.
Dissertation Defense Board:
Prof. Dr. Remigijus Paulavičius – Chair of the Council (Vilnius University)
Assoc. Prof. Dr. Viktor Medvedev (Vilnius University)
Assoc. Prof. Dr. Pavel Stefanovič (Vilnius Gediminas Technical University)
Assoc. Prof. Dr. Stepas Toliautas (Vilnius University)
Dr. Paulina Trybek (University of Silesia in Katowice, Poland)
21 September 2026
