Project start: 1.10.2025.

Computed Tomography (CT) is a technique for generating internal 3D images of the body. By providing detailed representations of bones and tissues, CT scans enable disease diagnosis, injury detection, surgical planning, and patient recovery monitoring. With its development, artificial intelligence has significantly transformed medical imaging diagnostics. Research in this field increasingly relies on deep learning to test hypothesis that artificial intelligence can improve targeted disease diagnostics and the development of new diagnostic software.

One of the common tasks associated with CT scans is the segmentation of organs and specific regions of interest, meaning assigning an exact label to each voxel in the scanned CT volume. Segmentation enables an in-detail examination of relationships between organs, their positioning, and the identification of anomalies important for further patient treatment. One of the most popular models in the segmentation of medical volumetric data, including CT, is the deep neural network nnUNet, known for its strong adaptability to input data. Despite the method's high adaptability, the direct impact of hyperparameters such as augmentation, normalization, input data size, neural network complexity, and model training settings on the final segmentation outcome remains insufficiently explored.

Therefore, this project will examine the correlation and influence of different hyperparameters on the final performance of deep learning models on selected publicly available datasets for CT image segmentation tasks. Additionally, the impact of the best-performing parameters will also be tested on clinical datasets from partner institutions (MedUni and HMS). To enhance practicality and education, a platform will be developed to support the training of neural networks for medical CT image segmentation. The project has the potential to improve segmentation techniques in medicine, contributing to the development of advanced medical diagnostic tools.