Given the significant security challenges posed to microservice systems in cloud computing, efficient anomaly detection is crucial for maintaining reliable microservice systems. Most existing anomaly detection methods require significant amounts of accurately labeled monitoring data or pure normal data. However, in the real-world scenario, the volume of labeled data is usually limited due to high cost, and the data often contains mislabeling noises owing to false alarms. To address the above problem, we propose MTMixAD, a microservice anomaly detection framework with limited and mislabeled monitoring data. Specially, MTMixAD uses Mixup for data augmentation, confident learning to filter mislabeled data, and a novel loss function to enhance model discrimination. Additionally, MTMixAD integrates metrics and trace data from microservice systems to better capture system characteristics. Experiments verify the effectiveness of our method and each part of MTMixAD.