Radiologists are increasingly relying on AI-based computer vision models to assist with the time-consuming task of interpreting medical scans. However, these AI models require vast amounts of accurately labeled data to function effectively, meaning radiologists must still invest significant time annotating medical images. To address this challenge, an international team led by Johns Hopkins Bloomberg Distinguished Professor Alan Yuille has developed AbdomenAtlas, the largest abdominal CT dataset to date. With over 45,000 3D CT scans and 142 annotated anatomical structures from 145 hospitals worldwide, AbdomenAtlas is more than 36 times larger than its nearest competitor, TotalSegmentator V2. This remarkable dataset and its findings were published in Medical Image Analysis.

Historically, abdominal organ datasets were created through the labor-intensive process of having radiologists manually label each individual organ in CT scans. This process required thousands of hours of expert labor. “Annotating 45,000 CT scans with 6 million anatomical shapes would require an expert radiologist to have started working around 420 BCE—the era of Hippocrates—to complete the task by 2025,” explains lead author Zongwei Zhou, an assistant research scientist at Johns Hopkins University.

To overcome this monumental task, the research team utilized AI algorithms to dramatically accelerate the annotation process. Collaborating with 12 expert radiologists and medical trainees, the team completed this ambitious project in under two years—something that would have taken humans alone more than two millennia.

The team’s solution involved a combination of three AI models, trained on publicly available labeled abdominal scans, to predict annotations for unlabeled scans. Color-coded attention maps were then used to highlight areas that required further refinement. This approach helped the team identify key areas for manual review by radiologists. By repeating this AI-prediction and human-review cycle, the researchers achieved a 10-fold speedup for tumor annotation and a 500-fold speedup for organ labeling.

The result was the creation of AbdomenAtlas—the largest fully annotated abdominal organ dataset in existence. This dataset continues to grow, with new scans, organs, and both real and artificial tumors being added. These contributions will help train AI models to detect cancerous growths, diagnose diseases, and even generate digital twins of patients.

“By enabling AI models to learn more about related anatomical structures before training on data-limited domains—such as in tumor identification—we have made AI perform similarly to the average radiologists in some tumor detection tasks,” says first author Wenxuan Li, a graduate student in computer science at Johns Hopkins.

AbdomenAtlas not only serves as a training ground for AI models but also as a benchmark to assess the accuracy of medical segmentation algorithms. The more data that can be tested, the more reliable and robust these algorithms become in complex clinical environments.

Looking ahead, the team plans to release AbdomenAtlas to the public, encouraging further advancements in medical segmentation challenges. At the 27th International Conference on Medical Image Computing and Computer Assisted Intervention last October, the team hosted the BodyMaps challenge, aimed at encouraging AI algorithms that are not only accurate but also efficient and practical in clinical settings.

Despite the groundbreaking nature of AbdomenAtlas, its creators note that the dataset represents only 0.05% of the CT scans acquired annually in the United States. To truly advance the field, the team calls for greater collaboration across institutions to increase data sharing, annotation, and the development of AI models.

“Cross-institutional collaboration is crucial for accelerating data sharing, annotation, and AI development,” the researchers conclude. “We hope that AbdomenAtlas will pave the way for larger-scale clinical trials and provide exceptional opportunities for practitioners in the medical imaging community.”

By Impact Lab