KSSD2025: A New Annotated Dataset for Automatic Kidney Stone Segmentation and Evaluation with Modified U-Net Based Deep Learning Models

Kidney stone disease is a prevalent condition associated with an increased risk of chronic kidney failure. Non-contrast Computed Tomography (CT) remains the gold standard for diagnosis due to its high sensitivity and ability to evaluate clinically relevant stone parameters. However, manual segmentation of stones in CT images is time-consuming and subject to inter-observer variability, highlighting the need for automated solutions. This paper introduces the KSSD2025, a novel dataset comprising 838 manually annotated axial CT images for kidney stone segmentation. We conducted a comprehensive evaluation of four segmentation models based on modified U-Net architectures: U-Net, U-Net++, U-Net3+, and TransU-Net. We evaluated their performance through 5-fold cross-validation and external tests. All models achieved mean Dice scores above 95%, with modified U-Net reaching 97.06% during cross-validation and modified U-Net3+ showing better generalization to external data. Our results confirm that U-Net-based architectures, when trained on high-quality annotations, can achieve accurate and reliable kidney stone segmentation. Our findings suggest that the KSSD2025 dataset offers a valuable resource to advance research in kidney stone analysis and the deployment of AI-assisted diagnostic tools in clinical practice.