A hybrid radiomics–3D CNN model using non-contrast CT for esophagogastric variceal bleeding risk prediction in cirrhosis
This study developed and validated a hybrid radiomics–3D CNN model using non-contrast CT that effectively predicts esophagogastric variceal bleeding risk in cirrhotic patients, offering a promising non-invasive tool for guiding prophylactic treatment decisions.
Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer
Liver cirrhosis is a condition where the liver becomes scarred and stiff, a state that forces blood to find new, dangerous pathways through the body. One of the most severe consequences of this pressure is the swelling of veins in the esophagus and stomach, known as varices. These swollen veins are fragile and can burst without warning, causing a life-threatening bleed that kills a significant portion of those affected. Currently, the standard way to check for this danger is an endoscopy, a procedure where a camera is passed down the throat. While effective, this method is invasive, uncomfortable for the patient, and difficult to repeat frequently, especially for those with other health complications. Doctors need a way to spot the risk early and often without putting patients through repeated invasive exams or exposing them to the risks of contrast dyes used in standard scans.
A team of researchers at Handan Central Hospital has explored a different path, using a type of medical imaging that is already common and safe: a non-contrast computed tomography, or CT, scan of the abdomen. Unlike the standard scans that use iodine-based dye to make blood vessels glow, this technique relies on the natural density of tissues. While these plain scans cannot show the swollen veins directly, the researchers suspected that the subtle changes in the liver, spleen, and surrounding blood vessels might hold the clues needed to predict a bleed. To unlock these hidden patterns, they combined two powerful ways of analyzing images: one that measures specific textures and shapes, and another that uses artificial intelligence to learn complex patterns on its own.
The study focused on 137 patients with liver cirrhosis who had undergone these plain CT scans. The researchers divided the group into two categories based on what happened over the next year: 47 patients experienced a variceal bleed, while 90 did not. The team taught a computer to look at seven specific areas in the scan, including the liver, spleen, esophagus, stomach, and three major veins. They used a method called radiomics to break down the images into thousands of tiny, measurable details, such as how rough the texture of the spleen appeared or the exact shape of the liver's surface. Simultaneously, they used a three-dimensional artificial intelligence model, a type of deep learning network, to scan the same images and learn its own set of features that a human might miss.
The researchers then built a "fusion model" that combined the detailed measurements from the radiomics approach with the complex patterns learned by the artificial intelligence. They tested this combined system by training it on most of the patients and checking its accuracy on the rest, repeating this process five times to ensure the results were reliable. The result was a tool that could distinguish between patients who would bleed and those who would not with a high degree of performance. The model achieved an AUC of 0.88. More importantly, it was very good at ruling out danger; the model achieved an AUC of 0.94 for identifying low-risk patients. This high level of certainty in identifying safe patients is crucial, as it could allow doctors to avoid unnecessary invasive procedures for those who do not need them.
The study also looked at how well the model's predictions matched reality. After adjusting the numbers to ensure they were trustworthy, the model showed that its probability estimates were highly reliable. When the researchers tested whether using this tool would actually help doctors make better decisions, they found that it offered a clear benefit over simply treating everyone or treating no one. The model performed better than using either the texture measurements or the artificial intelligence alone, suggesting that the two methods provide different, complementary information. The texture measurements captured the known physical changes in the organs, while the artificial intelligence found subtle spatial patterns that the human eye or simple measurements could not easily describe.
Despite these promising results, the researchers are careful to note that this is a pilot study with a relatively small group of patients from a single hospital. The model was not tested against standard clinical scores that doctors use today, nor was it tested on a larger, diverse group of people to see if it works everywhere. The process of outlining the specific body parts for the computer to analyze was done manually by doctors, which takes time and can vary slightly from one person to another. The authors suggest that while this approach shows great potential as a safe, repeatable, and non-invasive way to screen for bleeding risk, larger studies are needed before it can be used routinely in clinics. For now, the work demonstrates that even a plain CT scan, when viewed through the lens of advanced data analysis, can reveal critical information about a patient's future health.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.