A decentralized federated learning privacy protection method for medical scenarios was proposed to address the issues of slow model training convergence， gradient information leakage， single point of failure and vulnerability to attacks of central servers in most existing federated learning schemes in medical industry， resulting in problem of low global model robustness. Firstly， an efficient federated learning method based on Fletcher-Reeves （FR） conjugate gradient algorithm was constructed to achieve fast convergence， secure and reliable computation during the model training process. And the Paillier homomorphic encryption algorithm was used to solve the problem of gradient data leakage in medical institutions. Secondly， a decentralized trusted federated learning architecture based on blockchain consensus mechanism was proposed， thereby achieving secure modeling without the coordination of a trusted central server， and avoiding the problem of low training efficiency of the central server caused by attacks or paralysis. At the same time， global and local model updates were stored in the blockchain to achieve full lifecycle protection of the model. Security analysis results indicate that the proposed method possesses indistinguishability of ciphertext and data privacy. Simulation results show the proposed method has the advantages of fast convergence and high accuracy compared to the comparison methods， and can meet the practical needs of the existing medical data joint modeling better.