A two-stage u-net based hierarchical pipeline for kidney and renal tumor segmentation

Abstract

This report develops a two-stage 2D U-Net framework for automatic kidney and renal tumor segmentation from CT scans using the KiTS23 dataset. The first stage segments the kidney from full CT slices, while the second identifies tumors within a kidney-centered cropped region, reducing background noise and class imbalance. Five ImageNet-pretrained encoder architectures—ResNet-18, ResNet-50, DenseNet-121, EfficientNet-B3, and MobileNetV2—were evaluated under identical conditions. EfficientNet-B3 achieved the best kidney segmentation performance, with an F1 score of 0.9541 and IoU of 0.9122, whereas ResNet-18 performed best for tumor segmentation, achieving an F1 score of 0.7195 and IoU of 0.5619. The findings indicate that encoder performance depends on the specific segmentation task and that a hierarchical 2D approach can provide strong results with lower computational requirements than 3D models.