Cascade UNets for Kidney and Kidney Tumor Segmentation

Kidney and kidney tumor delineation constitutes an essential and labor-intensive task performed by expert radiologists and clinicians in order to diagnose renal-related pathologies, optimize radiotherapy treatment and speedup surgical planning. In this work, we address the problem of kidney delineation using 3D UNets to automatically segment the kidneys and kidney tumors and cysts in Computed Tomography (CT) series. A 3-stage cascade approach of UNets was applied to first segment kidneys and masses at low resolution with a coarse network, followed by a fine segmentation step at the second stage. During this refinement step, the predicted labelmap from the first stage is used as shape prior knowledge in order to guide the training of the model. At last, two independent UNets segment separately masses and tumors using the prior knowledge of the refinement step. Post-processing operations consist of 3D connected component analysis in order to remove false positive voxels out of the final predicted labelmap. We train our networks and evaluate the effectiveness of this approach in the KiTS23 challenge and dataset.