AAsU-Net: adaptive anisotropic convolutional network for renal neoplasm segmentation
Accurate segmentation of kidneys and renal tumors in computed tomography (CT) images is important for renal cancer diagnosis, treatment planning, and quantitative image analysis. However, abdominal CT volumes often show anisotropic resolution, with lower resolution along the slice direction than within the axial plane. This characteristic may reduce the effectiveness of conventional 3D convolutions and make it difficult to accurately segment tumors and boundary regions. In this study, we propose Adaptive anisotropic convolutional U-Net (AAsU-Net), an anisotropy-aware 3D segmentation network for renal neoplasm segmentation. The network includes an adaptive anisotropic convolution (AAs-conv) module, which combines spatially separable convolution and standard 3D convolution in parallel and adaptively fuses their features to improve representation under anisotropic imaging conditions. We also introduce a cross-scale feature fusion (CSFF) module in the encoder to preserve shallow structural details and reduce information loss during repeated downsampling. We evaluated the proposed method on the KiTS19 and KiTS21 datasets. On KiTS19, AAsU-Net achieved Dice scores of 0.970 ± 0.018 for kidney segmentation and 0.851 ± 0.031 for tumor segmentation, improving upon the baseline 3D U-Net by 1.19% and 3.30%, respectively. On KiTS21, the proposed method achieved a tumor Dice score of 0.879 ± 0.039 and an HD95 of 7.93 ± 2.74 mm, outperforming several recent methods, including 3D UX-Net, nnFormer, and SegMamba. These results suggest that AAsU-Net can improve segmentation accuracy and boundary delineation in anisotropic CT images and may provide useful support for automated renal tumor analysis.