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

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Standard Swin Transformers have limitations in modeling complex, non-linear vascular geometries due to fixed activation functions.
• Accurate segmentation of intricate vascular topologies in OCTA is crucial for clinical analysis.

Purpose of the Study:

• To introduce DDU-Net, a novel deep learning model for segmenting complex vascular structures in OCTA.
• To address the limitations of existing models in capturing non-linear vascular geometries.

Main Methods:

• Proposed KAN-Swin Transformer encoder with adaptive Kolmogorov-Arnold Networks (KANs) and B-spline activation functions.
• Developed a dual-path Double Dynamic Upsampler Module (DDUM) for parallel processing of shallow and deep features.
• Introduced an Information Compensation Module (ICM) and a multimodal fusion strategy with a Feature Alignment Module (FAM) for enhanced Inner Vascular Complex (IVC) segmentation.

Main Results:

• DDU-Net achieved state-of-the-art performance across five public datasets.
• Demonstrated statistically significant improvements through Wilcoxon signed-rank tests.
• Showcased enhanced capability in segmenting complex morphologies like bifurcations and high-tortuosity segments.

Conclusions:

• DDU-Net establishes a new reliable baseline for quantitative clinical analysis of OCTA images.
• The model's adaptive nature and novel modules effectively handle challenging vascular segmentation tasks.
• The proposed methods offer significant advancements in medical image analysis for vascular diseases.