Answers are grounded in this paper and cite their sources.
Medical image segmentation presents inherent challenges due to its complexity and heterogeneity. In this research, we propose an evolving segmentation model with dynamic transformer blocks and Gaussian, Levy, and 3D geometric surface-enhanced self-attention methods to extract fine-grained nuances and global contexts simultaneously. These new self-attention methods generate attention biases using Gaussian and Levy distributions and 3D formulae with patch distances as inputs to emphasize local regions. A new PSO variant is proposed to optimize the number of transformer blocks, self-attention variant types, encoder fusion weights, and key learning settings to fine-tune local and global dependency extraction to better adapt to different application contexts. The proposed PSO model employs sine-cosine search coefficients, reinforcement learning-based search action selection, and multiple optimal signals to avoid being trapped in local optima. Our model with the proposed PSO-optimized transformer architecture and learning settings achieves impressive mean Dice scores of 89.24%, 87.38%, 98.13%, 83.27%, and 82.73% for ACDC, CHAOS (MRI), CHAOS (CT), CHASE_DB1, and STARE, respectively. It also outperforms various recent state-of-the-art transformer and CNN baselines in tackling diverse binary and multi-class segmentation tasks.
No takes yet. Share an insight, caveat, or question.
Chen et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: