Accurate segmentation of multiple abdominal organs and co-occurring tumors in CT imaging presents a significant challenge due to high inter-patient anatomical variability, low contrast boundaries, and class imbalance. This paper proposes HybridSegNet, a novel architecture that integrates a Swin Transformer encoder with a multi-scale convolutional decoder equipped with dense skip connections and a dual-branch feature fusion module. HybridSegNet is evaluated on the publicly available CHAOS dataset (liver, kidney segmentation) and the KiTS23 challenge dataset (kidney tumor segmentation). The model achieves a mean Dice Similarity Coefficient (DSC) of 0.913 on liver segmentation, 0.897 on kidney segmentation, and 0.884 on kidney tumor segmentation, outperforming leading methods including nnU-Net (DSC: 0.876) and Swin-UNet (DSC: 0.861). A lightweight model variant is also proposed for deployment on resource-constrained clinical workstations without significant performance degradation. Results demonstrate HybridSegNet's suitability for real-time clinical decision support in abdominal oncology.
Jagadeesh et al. (Thu,) studied this question.