Supervised learning datasets often suffer from the problem of noise labeling, noise label filtering (NLF) has become an appealing topic. However, the phenomenon of class imbalance is easily neglected by recent NLF methods, which results in the dataset generated by these methods still limiting the model's generalization ability. A novel noise label filtering method called Prior Confidence Heterogeneous Consensus (PCHC) was proposed in this paper for class-imbalanced datasets. PCHC can improve the performance of the classifier by avoiding the aggravation of class imbalance caused by NLF. Firstly, prior confidence (PC) strategy is proposed to preliminarily divide the original dataset into clean dataset and noisy dataset. This strategy protects minority-class samples by designing new label confidence metrics and increasing the confidence threshold for minority class samples to be recognized as noise samples; Secondly, heterogeneous consensus (HC) strategy is proposed to mine potential minority-class pseudo samples in noisy datasets. Two heterogeneous classifiers are trained on clean dataset in this strategy and the fusion decision-making ability of sample characteristics and sample space features used for mining minority-class pseudo samples in noisy dataset. Experiments on four publicly available imbalanced financial datasets have shown that the dataset generated by the PCHC method is more conducive to improving the model's generalization performance than the dataset obtained by the baseline method.