Whole body imaging is a technique that scans a person’s insides for the early warning signs of heart disease, cancer and other worrying conditions. There are various ways to make these scans but the most common uses x-ray to create images of body slices. A computer then fits the images back together to create a 3D model of the whole body.
This can be used to plan certain types of surgery but it is also offered as a kind of screening service to provide piece of mind to health-conscious individuals—at least that’s the promise.
The reality is that whole body CT scans are difficult to analyze, not least because it is hard to identify all the different organs from the mass of tissues that make up the human body. So physicians have turned to computer vision systems to do the job instead.
The task is to identify the organs, structure and bones, as well as their three-dimensional shape using the data from the scan. However, the current crop of algorithms do not work particularly well, say Jakob Wasserthal and colleagues at the University Hospital Basel in Switzerland.
Noisy Data
Some of these algorithms are hard to use, others have been trained on ultra-clean images and so do not work on the noisier images that radiologists take in a clinical setting.
Another problem is that one algorithm might only work for a single anatomical structure or for a small number of them. This forces physicians to use several algorithms for different structures, which dramatically increases the complexity of the data analysis process.
What’s needed, say Wasserthal and co, is an easy-to-use algorithm that can accurately segment multiple organs and body structures using real world scans.
As it turns out, the team have developed an algorithm, called TotalSegmentator, that does exactly this. In tests, it works well. “In 1204 CT images, we segmented 104 anatomical structures (27 organs, 59 bones, 10 muscles, 8 vessels) covering a majority of relevant classes for most use cases,” say the team. pointing out that a human annotator would 10 about ten years to perform an equivalent task manually.
The algorithm isn’t perfect, however. For example, it sometimes misses small sections of tubular structures, such as arteries and the colon. It can also mix up classes of neighboring vertebrae and ribs. “Being aware of those can be helpful when using the segmentations of the affected classes,” say Wasserthal and co.
In other cases the algorithm is highly accurate. The team put the algorithm through its paces by analyzing how body structures change throughout a human lifespan. Using a dataset of over 4000 whole body CT scans of people ranging in age from 18 to 100 years, they asked the algorithm to automatically identify structures, such as the liver and left atrium of the heart. They cold then see how these structures changed over time.
They found that the volume and density of the left atrium increases throughout life. By comparison, the liver increases in volume up to the age of 50 and then shrinks. “The same analysis can easily be done for all of the 104 classes [of body structures]” say the team.
That’s impressive work that could make whole body scans more useful and more common. Whether they will put anybody’s mind at rest is another question entirely. The US Food and Drug Administration says it “knows of no scientific evidence demonstrating that whole-body scanning of individuals without symptoms provides more benefit than harm to people being screened.”
More evidence is obviously needed. Perhaps TotalSegmentator can help.
Ref: TotalSegmentator: robust segmentation of 104 anatomical structures in CT images : arxiv.org/abs/2208.05868
