To improve the performance of panoramic segmentation of complex maxillofacial structures, a multi-scale feature extraction model (MC-UKAN) was proposed, which combined the Kolmogorov-Arnold representation theorem with the U-Net architecture. The model integrated position priors, learnable nonlinear activation functions, and multi-scale feature extraction to effectively enhance feature expression ability. Based on this model, a three-stage maxillofacial structure panoramic segmentation framework was designed. In the first stage, coarse segmentation and tooth position calibration were performed on low-resolution images using MC-UKAN. In the second stage, the target structures were localized using the coarse segmentation results on the original images, the K-means clustering algorithm was then applied to categorize these targets into five classes, and corresponding networks were employed for fine-grained segmentation. In the third stage, a lightweight network was used to segment the jawbones and pharynx on the original images. By fusing multi-scale segmentation results, precise segmentation of structures such as teeth, dental pulp, jawbones, and nerve canals was achieved. Experimental results on the ToothFairy3 dataset demonstrated that the proposed method achieved an 88.3% Dice coefficient and 5.04 mm HD95 for the segmentation of 74 categories of oral and maxillofacial structures, with an average inference time of 27.04 seconds. These results fully validated the superior performance of the proposed method in complex maxillofacial structure segmentation tasks.