Dynamic Resolution Network for Kidney Tumor Segmentation

Segmentation of kidneys, kidney tumors and kidney cysts from contrast-enhanced CT images has significant potential to facilitate large-scale imaging and radiological analysis. However, the task is challenging due to the considerable variation in tumor scales between different cases, which is not effectively addressed by conventional segmentation methods. In this paper, we propose a method called dynamic resolution that addresses this issue by dynamically adjusting the image resolution for each sample during training and testing, thus achieving a balance between targets with different scales. We also present a technique that uses publicly available unlabelled datasets to improve the robustness of the model without requiring additional manual labelling. We evaluated our method on the KiTS23 competition dataset and the results demonstrate its superiority over the existing state-of-the-art nnUnet, with improvements of 1.2%, 3.9% and 4.9% on kidney, tumor+cyst and tumor respectively.