Abstract: Tri-training is a promising semi-supervised learning approach for Computer Aided Diagnosis (CAD) systems. It aims at enhancing the performance of the hypothesis that is learned on only a small amount of expert-diagnosed samples by utilizing the large amount of undiagnosed samples through co-training process. However, mislabeling the unlabeled samples in the co-training process is inevitable and harms the performance improvement of the hypothesis. In this study, we extend the co-training process by a participatory learning cognition paradigm and propose a new tri-training algorithm named PL-Tri-training. In detail, the acceptance unit of participatory learning is instantiated as a data editing operation and the critic unit of participatory learning is designed as an adaptive arousal strategy for the data editing. In the co-training process of PL-Tri-training, the acceptance unit utilizes data editing to identify and remove the mislabeled data, as well as the critic unit exploits arousal strategy to inhibit the invalid activation of data editing. Experiments on three benchmark medical data sets verify the effectiveness of the proposed algorithm. A successful application to the pulmonary nodules detection in chest CT images shows that PL-Tri-training can more effectively exploit the undiagnosed samples to improve the diagnosis performance than Tri-training and AC-Tri-training, which extends the co-training process only with the acceptance unit of participatory learning.