CNN U-Net

Manual 3D segmentation of Wilms’ tumor and kidneys is highly time-consuming, limiting routine clinical use, prompting this study to develop an AI-assisted segmentation approach for pediatric CT scans. While fully automated U-Net segmentation performed poorly, the proposed OV2ASSION training strategy—combining sparse manual input with CNN-based segmentation—achieved expert-level accuracy with Dice scores up to 0.97 for tumors and 0.94 for kidneys. This hybrid approach reduced expert intervention time by approximately 80%, demonstrating a practical and accurate solution for clinical 3D reconstruction in children with Wilms’ tumor.