NIH Releases Huge Database of CT Images for AI Testing

The National Institutes of Health’s {(NIH) Bethesda, MA, USA} Clinical Center has made a large-scale dataset of CT images available to the public in order to help the scientific community improve the detection accuracy of lesions. The dataset, named DeepLesion, has over 32,000 annotated lesions identified in CT images, as compared to less than a thousand lesions in most of the publicly available medical image datasets. The images are of 4,400 unique patients, who are partners in research at the NIH and have been completely anonymized. In 2017, the NIH clinical center had released anonymized chest X-ray images and their corresponding data.

The NIH Clinical Center is the clinical research hospital for the NIH, the US’ medical research agency that includes 27 institutes and centers and is a component of the US Department of Health and Human Services. NIH is the primary federal agency conducting and supporting basic, clinical, and translational medical research, and is investigating the causes, treatments, and cures for both common and rare diseases.

Radiologists at the clinical center use an electronic bookmark tool to measure and mark clinically meaningful findings from the CT images of patients. The radiologists save their place and mark significant findings, which they can visit again at a later time. These complex bookmarks provide arrows, lines, diameters, and text that can pinpoint the precise location and size of a lesion to allow experts to identify growth or a new disease.

Scientists at the NIH clinical center have used these bookmarks, which are abundant with retrospective medical data to develop the DeepLesion dataset. Unlike most lesion medical image datasets that are currently available that can only detect one type of lesion, DeepLesion offers greater diversity as it contains all types of critical radiology findings from all over the body, such as lung nodules, liver tumors, enlarged lymph nodes, and so on. The dataset released by the NIH is large enough to train a deep neural network and could enable the scientific community to create a large-scale universal lesion detector with one unified framework.

Researchers hope that by making the medical image datasets publicly available, others will be able to develop a universal lesion detector that will help radiologists identify all kinds of lesions. It can also serve as an initial screening tool and send its detection results to other specialist systems trained in certain types of lesions. DeepLesion could also help radiologists to mine and study the relationship between different types of lesions in order to make new discoveries. It can allow them to more accurately and automatically measure the sizes of all lesions in a patient, thus enabling full body assessment of cancer.

The NIH clinical center plans to continue improving the DeepLesion dataset by collecting more data and further increase its detection accuracy. Its universal lesion detecting capability will become more reliable after researchers manage to leverage 3D and lesion type information. In future, the application of DeepLesion could be further extended to other image modalities such as MRI and combined with data from various hospitals.

Related Links:
National Institutes of Health