The application of Magnetic Resonance Imaging (MRI) technology has provided key technical support for the precise segmentation of the Anterior Visual Pathway (AVP), which is of great significance for analyzing its X-shaped three-dimensional morphological characteristics and neural trajectories. Despite the remarkable achievements made in previous studies, three main challenges remain: loss of fine anatomical structural features, insufficient retention of tissue texture details, and discontinuity of segmentation results. The root causes of these challenges are the insufficient integration of multi-scale features and the limited ability to model global context. To address the above challenges, our study proposes the Multi-scale Deform Kolmogorov-Arnold Networks (MD-KAN). This architecture innovatively integrates three core components: the Multi-scale Feature Fusion (MFF) module, the Deform Attention module, and the Vision KAN module. Specifically, the MFF module adopts a dual-path architecture, which adaptively fuses multi-scale spatial information by processing feature maps with different scales in parallel. The Deform Attention module effectively simulates the local self-attention process through a dynamic convolutional kernel offset mechanism, thereby accurately modeling the complex curvature features and local geometric shapes of the AVP. The Vision KAN module employs a parametric basis function combination strategy and a hierarchical feature interaction mechanism, which enhances the non-linear expression capability while effectively integrating global context information. Extensive experiments conducted on the AVP-119 and AVP-83 datasets demonstrate that the MD-KAN network outperforms other state-of-the-art medical image segmentation networks in terms of performance.