DDU-Net: learning complex vascular topologies with KAN-Swin transformers and double dynamic upsampler

To segment complex vascular topologies in Optical Coherence Tomography Angiography (OCTA), we introduce DDU-Net. This work addresses the theoretical limitations of standard Swin Transformers, whose internal Multi-Layer Perceptron (MLP) blocks use fixed activation functions, restricting their capacity to model non-linear vascular geometries. We propose the KAN-Swin Transformer, an encoder block that replaces this rigid component with an adaptive operator based on Kolmogorov-Arnold Networks (KANs). This new layer features B-spline-based learnable activation functions on network edges, rather than fixed functions on nodes, empowering the encoder to learn geometrically-aware representations specific to intricate morphologies like bifurcations and high-tortuosity segments. The decoder features a novel dual-path Double Dynamic Upsampler Module (DDUM), which processes edge-rich shallow features and semantic deep features in parallel before an attention-based fusion, avoiding feature contamination. An Information Compensation Module (ICM) further recovers fine details using multi-dilation convolutions. For challenging low-contrast Inner Vascular Complex (IVC) images, we introduce a multimodal fusion strategy, where a Feature Alignment Module (FAM) aligns probability maps from auxiliary modalities to enhance the IVC representation. Extensive experiments on five public datasets demonstrate that DDU-Net achieves state-of-the-art performance. Rigorous Wilcoxon signed-rank tests confirm these improvements are statistically significant, establishing DDU-Net as a reliable new baseline for quantitative clinical analysis. The code for DDU-Net is available on GitHub (https://github.com/steve706/DDUM_final).