Renal tumor, along with renal cyst, is one of the most common kidney diseases. As the kidney tumor incidence is increasing, there is a need for efficient diagnosis and reliable treatment outcomes predictions. Automatic kidney images characterization and differentiating between tumors and cysts could help clinicians with these procedures, providing rapid and repeatable results, free from interobserver variability. The aim of this study is to develop a model for segmentation of kidneys, kidney tumors and cysts on CT scans. For this task we employ a transformer based architecture - Swin UNETR. We conducted a series of experiments to determine which hyperparameters improve the overall model performance measured by Dice score and the model metrics for each class separately. Our best performing model achieves the following Dice scores on test dataset: overall: 51.3, kidney + masses: 78.8, masses: 39.6, tumor: 35.5 and the following Surface Dice scores: overall: 22.6, kidney + masses: 36.7, masses: 16.7, tumor: 14.5. Our model ranked 24th on the leaderboard. The code for our solution is publicly available at https://github.com/deepdrivepl/kits23 .