Leveraging Uncertainty Estimation for Segmentation of Kidney, Kidney Tumor and Kidney Cysts

In the field of medical imaging, computed tomography (CT) scans have become crucial for the detection and management of anatomical abnormalities. This study presents an improved cascaded nnUNet framework incorporating a cropping strategy and uncertainty estimation for effective segmentation of kidneys, kidney tumors, and kidney cysts in computed tomography scans. The proposed method is evaluated on the KiTS23 dataset, consisting of 489 CT scans with accompanying masks for the kidney, tumor, and cyst. We exploited a low-resolution nnUNet for initial kidney segmentation, and the resulting predictions were used to crop a bounding box area to decrease data dimensionality, which facilitated faster training and inference. A cyclic learning rate was applied along with posterior sampling of the weight space, enabling an ensemble of five models from different training cycles. This approach showed superior performance, particularly in the segmentation of tumors and masses, as compared to other models such as the standard nnUNet, the cascaded nnUNet, and the BANet. Moreover, our ensemble model, including models from different training cycles, indicated a strong correlation between predicted uncertainty maps and false positive detection, holding promising potential for enhanced clinical utility.