Self-Supervised Kidney Tumor Segmentation Using Random Block Reconstruction on 3D CT Scans

Background : Self-supervised learning (SSL) has shown strong potential to improve medical image segmentation by leveraging unlabeled data for representation learning. However, most existing SSL methods are designed for 2D natural images and rely heavily on semantic transformations that do not generalize well to volumetric medical imaging. To address this limitation, we introduce Random Block Reconstruction (RBK), a novel self-supervised pretraining strategy tailored for 3D CT scans. RBK focuses on reconstructing randomly removed spatial blocks within the input volume, promoting the learning of spatially consistent and domain-relevant representations. Methods : We compare RBK with two established SSL methods, Masked Generative modeling (MG) and Bootstrap Your Own Latent (BYOL), on the KiTS23 kidney tumor segmentation dataset. All methods were pretrained on 80% of the dataset and fine-tuned on the remaining 10%, using nnU-Net as the segmentation backbone. Each cross-validation fold was repeated three times to mitigate randomness, with Dice Similarity Coefficients (DSC) and Normalized Surface Dice (NSD) metrics evaluated for kidney, cyst, and tumor segmentation. Results : RBK achieved the highest overall performance, improving the average Dice score by 2.82% and the Surface Dice by 3.9% over the supervised baseline. Notably, the largest gains were observed in tumor segmentation, where RBK outperformed both MG and BYOL by up to 4.7%. Despite using only 10% of the labeled data for fine-tuning, RBK reached segmentation accuracy comparable to state-of-the-art supervised methods trained on the full dataset, demonstrating efficient transfer of learned representations. Conclusion : The RBK pretraining approach shows that spatial reconstruction can effectively guide representation learning in volumetric medical imaging without relying on semantic transformations. Its consistent improvements across segmentation tasks emphasize its methodological robustness and potential for broader use in medical image analysis.