Font Size: 

Accurate segmentation of CT images containing kidneys, tumors and cysts is crucial for diagnosis and treatment planning. This study aims to reveal the impact of different deep learning architectures and backbone structures on class-based performance in the segmentation of kidneys, tumors, cysts in CT images, thereby establish a foundation for future research aimed at developing more effective segmentation architectures. To this end, U-Net, U-Net++, SwinUNETR, SAM2, YOLOv8-Seg, nnU-Net models were implemented. The models were evaluated quantitatively using the Dice Similarity Coefficient (Dice), Normalized Surface Dice (NSD), qualitatively through the resulting segmentation images. For kidney segmentation, nnU-Net demonstrated the highest performance, achieving a Dice score of 0.9579, an NSD of 0.9396. SAM2 achieved Dice, NSD values ​​of 0.8931, 0.9277 for tumor segmentation, and 0.7951, 0.8948 for cyst segmentation. The findings revealed that model performance varies depending on the target anatomical structure and that performance differences between models are particularly pronounced when segmenting small and variable structures.