DAMFF-Net: 3D renal tumor segmentation based on dynamic attention multi-scale feature fusion network

Abstract Accurate segmentation of renal tumors is crucial for the diagnosis and treatment of the urinary system. However, due to the low contrast between renal tumors and surrounding tissues, existing segmentation methods have certain limitations in dealing with fuzzy boundaries of renal tumors. Therefore, to address these issues, we propose a 3D U-shaped network based on dynamic attention multi-scale feature fusion (DAMFF-Net). The core idea of the proposed model is to fuse rich target multi-scale structural features in multidimensional medical images through dynamic attention guidance. Specifically, we introduce a recurrent slice-wise attention mechanism, which proficiently models the long-range dependencies among slices and augments the model's comprehension of global spatial features. Subsequently, we propose a dynamic attention multi-scale feature fusion module to elevate the model’s capacity to express multi-scale features of renal tumors. This module dynamically adjusts the weights of features across different scales through an attention weighting mechanism and achieves efficient multi-scale fusion by integrating dense multiplicative connections. The proposed network is validated on the KiTS23 dataset, and experimental results indicate that compared to the baseline model. The result shows improvements of 3.11%, 2.34%, 0.51%, 1.51% and 2.07% in IoU, dice, accuracy, precision and recall, respectively. The experimental results demonstrate that the algorithm outperforms other algorithms in both visual effects and objective evaluation metrics. Furthermore, test results based on the private dataset of the First Affiliated Hospital of Zhengzhou University show that DAMFF-Net still performs well in terms of generalization, which further validates the potential value of our model in clinical applications. The code is available at https://GitHub-ahuweia/DAMFF-Net.