In this study, we tried to create a machine learning method that detects disease lesions from CXR images using data set annotated with extracted CXR image reports information. We set nodule as the target disease lesion and used the report information to automatically produce training data for object detection task. It could be considered as semi supervised learning. Our project totally had four modules. Firstly, we use U-Net to segment the lung field according to report record. Secondly, we made the classification model of nodule which could discriminate nodule and non-nodule image in left and right lung with DenseNet-121. The area under ROC (AUC) was 0.771 and 0.788 for left and right lung respectively. Next, we also used attention map which visualized the outcome of computer prediction. Besides, Self Intersection (SI) technology was applied to find whether the attention area was consistent with report record of nodule. It could filter the attention area possibly with nodule and make teacher data for object detection model. In addition, some annotation of nodule in 457 chest x-ray images from test set were made by physician in order to evaluate the attention area generated by SI. We defined our new standard for evaluating the consistency between man made annotation and computer prediction. The results in test set showed 54.6% consistency accuracy for man made annotations and 73.2% consistency accuracy for computer prediction. Finally, we taught the object detection model Faster-RCNN with bounding boxes produced by SI. The outcome in test set improved the consistency of man made annotation to 60.7%. These results demonstrated that our proposed method could partially generated teacher data for object detection. The accuracy of the results may be improved by learning more images.