Our dataset consists of three parts: raw DICOM data, JPG images transformed from raw DICOM data, and non-image data including sex, age, history, some patients have gene expression, and pathologist reports. The images were analyzed both on the mediastinum (window width, 350 HU; level, 40 HU) and lung (window width, 1,400 HU; level, –700 HU) settings. The reconstructions were made in 2mm-slice-thick and lung settings. The CT slice interval varies from 0.625 mm to 5 mm, scanning mode include plain scan, contrast scan, 3D reconstruction, etc. All the cases were confirmed by pathological diagnosis. We labeled the locations of tumor in JPG images. And the image annotations are saved in XML files in PASCAL VOC in Annotation Files with Hashed Filenames format. Users can parse the annotations using the PASCAL Development Toolkit. We provide JPG images and XML annotation files in PASCAL VOC format which is widely used in deep learning and machine learning researches. The annotation files are provided by five doctors and two deep learning researchers. Besides that, all the cases were confirmed by pathology. Thus, we can guarantee our dataset precise and ease of use. Our dataset can be regarded as a useful tools and data resource to develop medical diagnosis algorithm based on deep learning. On the other hand, our data set can be used as an effective tool for promoting medical diagnosis. | 