Early diagnosis of thorax diseases may improve a patient's chances of cure and recovery. Recently, deep learning approaches are applied to multilabel classification of chest X-ray images. However, multilabel causes imbalance in the train data is a problem that happens with a variety of data, especially health data. This study aims to improve the performance of diseases detection from X-ray images. After adjusting the balance of data sample among different disease labels, a voting classification method has been involved to combine the results from different models. As a result, a meaningful improvement has been achieved.