In order to improve the performance of CT image's lesion recognition algorithm and improve the diagnosis accuracy of doctors, a CT lesion recognition algorithm based on improved particle reseeding method is proposed. First of all, aiming at the non- uniformity of topography, the Lagrangian labeled particles are calculated before the level set formula is calculated to reconstruct the embedded interface, thus improving the quality conservation characteristics of the level set algorithm. Secondly, in view of the uncertainty of the traditional particle method in dealing with interface singularity and complex geometry- related problems, the convergence of velocity fields at singular points and topological change points is promoted by adding velocity vectors and unit normal vectors. Finally, the effectiveness of the proposed algorithm is verified by simulation experiments.