Multilevel Feature Fusion and KAN Integration for Brain Tumor Segmentation

In recent years, computer-aided medical image diagnosis has gained significant traction, particularly in the realm of medical image segmentation, where deep learning techniques like U-Net have demonstrated powerful capabilities. Despite their effectiveness, conventional models often struggle with interpretability and the intricate, irregular structures found in medical images. This study introduces MFFKAN-UNet, a novel segmentation network that integrates the Kolmogorov-Arnold Network (KAN) to enhance interpretability and nonlinear modeling while reducing parameter complexity. Our approach employs a multi-level feature fusion mechanism enhanced by novel attention mechanisms, which improves the capturing of complex relationships between features. We further incorporate a lightweight, learnable up-sampling technique and a novel Boundary DoU Loss function to refine segmentation accuracy. Evaluated on the BraTS 2019 dataset, MFFKAN-UNet demonstrates superior performance compared to existing segmentation methodologies, highlighting its potential for advancing medical image analysis.