We propose the Diffusion-Based Conv-Former U-Net (DCFDU-Net) model for panoramic CT segmentation tasks. Our model primarily employs a dual-encode structure, comprising the CMT and PVT modules. To enhance boundary precision, we incorporated a novel boundary learning module inspired by DDPM and level set. This module constructs the level set function by initially predicting the boundary in a high-dimensional space projection. Then uses the DDPM model to evolve this projection, facilitating the accurate delineation of the zero level set. Finally, the boundaries and mask outcomes are refined through an efficient, cost-effective network architecture. Our method achieved an average DICE score of 91.81% and an average IOU score of 96.35%, with an average HD distance of 0.0332 for teeth segmentation on the validation set using an NVIDIA GeForce RTX 3090 GPU. The average running time was 0.91 s per image. The code is available at https://github.com/aoxipo/AITOOTH .