Published on
Updated
Reading 3 mins.
American researchers have developed a model based on Artificial Intelligence capable of predicting the occurrence of future lung cancer up to six years in advance, thanks to a single low-dose scanner.
The fight against lung cancer certainly promises to be technological. Known as one of the deadliest, this cancer is very difficult to detect early before symptoms appear. In February 2022, the High Authority for Health (HAS) recommended the implementation of a small-scale trial of lung cancer screening among smokers.
This screening is already in place in several countries around the world, such as the United States or Canada, for example. In France, the reluctance vis-à-vis the false positives and the “overdiagnosis” that the examination would generate in particular, as we already mentioned in our article “Lung cancer screening: for or against?”, have long slowed down the development of such a program.
In the United States, screening by low-dose scanner is already recommended and a team from the Massachusetts Institute of Technology (MIT) is now proposing an improvement with the use of artificial intelligence.
Lung cancer: facilitating screening with AI
The Massachusetts Institute of Technology team started from the observation: low-dose computed tomography (LDCT) for lung cancer screening is effective, but most eligible people are not screened. “We hypothesized that a deep learning model evaluating the entire LDCT volumetric data could be constructed to predict individual risk without requiring additional demographic or clinical data.” they explained in a study published on January 12 in the Journal of clinical oncology. Thus was born the Sybil model.
Deep learning or “deep learning” is a sub-branch of artificial intelligence, which aims to build knowledge from a large amount of information (the greater the amount, the more knowledge will be refined). In the case that interests us, it is a question of “feeding” an algorithm with numerous low dose scan images and of having in parallel the medical history of the patients who underwent this imaging examination (the occurrence or not of lung cancer).
A prediction deemed solid between 1 and 6 years after the screening
Current lung cancer prediction models require a combination of demographic information, clinical risk factors, and radiological annotations. Sybil is designed to use a single low-dose chest scan to predict the risk of lung cancer occurring 1-6 years after a screening.
The results are encouraging. Applied to a diverse set of analyzes from two hospitals and the National Lung Cancer Screening Trial, the study showed that Sybil was able to predict short- and long-term lung cancer risk, obtaining C index scores ranging from 0.75 to 0.80. Values above 0.8 indicate a strong model, the study says.
When predicting cancer risk a year in advance, the model was even more efficient: they scored between 0.86 and 0.94 on a probability curve.
Predict cancers yet invisible on scans
Still an astonishing indication in the field: the imaging data used to train Sybil was, however, largely free of any signs of cancer. There’s a reason, though: at an early stage, lung cancer occupies small parts of the lung – just a fraction of the hundreds of thousands of pixels that make up each CT scan. Co-author Jeremy Wohlwend was surprised by Sybil’s high score, despite having no visible cancer.
“We found that even though we humans couldn’t quite see where the cancer was, the model could still have some predictive power as to which lung would eventually develop cancer.”
Already undertaken in other cancers, as well as intestinal diseases or Alzheimer’s disease, Artificial Intelligence based on imaging data promises to be a tool allowing an unparalleled degree of precision to be achieved.