AI-Powered Abdomen Map Enables Early Cancer Detection

By MedImaging International staff writers
Posted on 05 Feb 2025

Image: Two abdominal CT scan slices, standard on the left and AbdomenAtlas’ organ segmentation on the right (Photo courtesy of Johns Hopkins University)

Radiologists are increasingly turning to artificial intelligence (AI)-driven computer vision models to assist with the time-consuming task of analyzing medical scans. However, these models require large, meticulously labeled datasets to produce reliable and accurate results, meaning radiologists still need to spend significant time annotating medical images. To address this challenge, researchers have harnessed AI to create the most extensive and comprehensive dataset of abdominal organs to date, designed to help radiologists quickly and accurately identify tumors and other medical conditions.

A team of researchers from around the globe, led by Johns Hopkins University (Baltimore, MD, USA), has developed AbdomenAtlas, which is the largest abdominal CT dataset available, consisting of more than 45,000 3D CT scans representing 142 annotated anatomical structures. These scans were gathered from 145 hospitals worldwide, making AbdomenAtlas more than 36 times larger than the nearest competitor, TotalSegmentator V2. The dataset and its applications were featured in a recent edition of Medical Image Analysis. Prior abdominal organ datasets were compiled by radiologists manually identifying and labeling individual organs in CT scans, a labor-intensive process requiring thousands of hours of human effort. The researchers accelerated this annotation process by leveraging AI algorithms. With the help of 12 expert radiologists and additional medical trainees, they completed in under two years a task that would have taken human annotators over two millennia.

The method developed by the researchers combines three AI models trained on publicly available datasets of labeled abdominal scans to predict annotations for previously unlabeled scans. The AI then generates color-coded attention maps that highlight areas needing refinement, allowing radiologists to focus their review on the most critical sections of the models' predictions. This iterative process—AI prediction followed by human validation—greatly accelerates the annotation workflow, achieving a 10-fold speedup for tumors and a 500-fold speedup for organ annotations, according to the researchers. This process not only increases the scope, scale, and precision of the dataset, but it also prevents overwhelming the radiologists involved. The outcome is what the team describes as the largest fully annotated abdominal organ dataset available. They are continuing to expand the dataset by adding more scans, organs, and both real and artificial tumors, further enhancing the training of AI models for identifying cancer, diagnosing diseases, and even creating digital twins of real-life patients.

AbdomenAtlas also provides a valuable benchmark that allows other research groups to test the accuracy of their medical segmentation algorithms. The more comprehensive the dataset used to evaluate these algorithms, the more reliable and effective the models can be in complex clinical scenarios, according to the Hopkins team. The researchers plan to make AbdomenAtlas publicly available and are introducing new medical segmentation challenges to encourage AI algorithms that are not only theoretically robust but also practically efficient and reliable for use in clinical settings. Despite the advancements offered by AbdomenAtlas, its creators note that it represents only 0.05% of the CT scans acquired annually in the United States, and they are calling on other institutions to help expand this important resource.

"Cross-institutional collaboration is crucial for accelerating data sharing, annotation, and AI development," the researchers wrote. "We hope our AbdomenAtlas can set the stage for larger-scale clinical trials and offer exceptional opportunities to practitioners in the medical imaging community."