Feature-aggregated attention U-Net for kidney and kidney tumor segmentation in computed tomography images

Abstract

Accurate computed tomography (CT)-based segmentation of kidneys and renal tumors provides quantitative anatomical information that can support downstream volumetric analysis and treatment-planning workflows. This study presents FAU-Net, a Feature-Aggregated Attention U-Net that combines Cross-Channel Attention (CCA), Multi-Layer Feature Fusion (MFF), and decoder-stage Spatial Attention (SAM) within a unified two-dimensional encoder-decoder framework. Patient-grouped partitioning of the KiTS19 cohort was used so that all axial slices from a given patient remained exclusively within the training, validation, or test subset. Model optimization used an equally weighted combination of weighted cross-entropy and Dice loss to provide class-reweighted pixel-wise supervision while directly optimizing segmentation overlap. FAU-Net achieved Dice scores of 0.9620 and 0.9421 for kidney and tumor segmentation, respectively, corresponding to a mean Dice of 0.9521. The respective IoU scores were 0.9451 and 0.9251, yielding a mean IoU of 0.9351. Architectural ablation demonstrated progressive gains after incorporating CCA, MFF, and SAM, supporting their complementary roles in feature recalibration, multi-scale aggregation, and spatial refinement. The reported results demonstrate strong segmentation performance on target-containing axial slices within the defined KiTS19 research protocol. At the same time, complete-volume and external validation provide complementary evaluation targets for broader clinical translation.