MedTxt-RR: Radiology Reports Corpus

There is no scarcity of research on the use of medical artificial intelligence for automatically generating  reports from medical images such as simple X-ray scans, CT scans, or MRI scans. But, there exists no standard method for evaluating the quality of these critical medical reports. The reason for this is that the same medical result from a clinical scan may be written stylistically differently by different doctors. And, conversely, reports with different medical content may be stylistically similar. Furthermore, as standard practice in the medical field, only one reading report is created for an image, so simply collecting reading reports stored in medical institutions does not solve this problem as we don’t get sufficient information on how different doctors would write the same medical content.

In order to gather information on how medical contents are expressed, we asked multiple doctors to write a reading report for the same CT image(*1). The reports were created under a single criteria of “judging the stages of lung cancer from CT images by TNM classification (*2)”. The dataset thus created will help understand how diversely the medical contents of are expressed on medical image readings. A sample of the dataset is available on this site under the name of “Public Image Reading Report Dataset”, so please feel free to use it.