Kidney Tumor Segmentation With a Multistage Adaptive Boundary-Aware Network

Accurate kidney tumor segmentation is critical for surgical planning but is challenged by indistinct boundaries and high morphological variability in computed tomography (CT) images. We propose the adaptive boundary-aware network (MABS-Net). The architecture integrates three core innovations: (1) a boundary-aware multiscale feature extraction module using learnable boundary-enhancing convolutions and adaptive weight maps to capture subtle edge cues; (2) an adaptive three-stage cascaded strategy for progressive refinement from coarse localization to uncertainty-driven boundary optimization; and (3) a contrastive learning mechanism with online hard example mining to explicitly boost feature discrimination between tumor and normal tissues in ambiguous regions. Experiments on the KiTS19 and KiTS21 datasets demonstrate MABS-Net's superiority. On KiTS19, it achieved a Dice coefficient of 0.891 ± 0.034, significantly outperforming the nnU-Net baseline. Notably, the 95% Hausdorff distance (HD95) was reduced to 6.73 ± 2.28 mm, and the boundary Dice score improved by 5.8% compared to state-of-the-art methods, validating our boundary-aware design. Furthermore, the model provides pixel-wise uncertainty maps to support reliable clinical decision-making. MABS-Net balances high accuracy with computational efficiency (0.53 s/case), presenting a promising solution for automated renal tumor analysis.