The leading cause of lung cancer is smoking, and early detection can save lives. However, lung cancer screening has historically been inconvenient, requiring expensive CT scans that expose patients to radiation. Researchers from the Massachusetts General Cancer Center and MIT have developed Sybil, a deep-learning tool to change this. Using a data set of more than 20,000 LDCT scans, Sybil predicts a patient’s lung cancer risk for the next six years. As a result, current and former smokers, as well as those who never smoked, will be able to get screened more easily for lung cancer.
Within a year of screening, Sybil had an AUC (area under the curve) value of 94%, which indicates that it was successful in correctly classifying people with or without lung cancer. The false positive rate for lung cancer has been reduced from 14% to 8% for the first scan, which opens up the possibility of accurate results in just one scan.
To assess Sybil’s performance in different ethnic groups, further evaluation is required.
In the United States, lung cancer is the second-most common type of cancer and a leading cause of death. The primary factor in its development is smoking cigarettes – however, people who quit can significantly reduce the risk. Screening proved to be effective in lowering the mortality rate to 20%
Right now, CT scans are widely used to screen for lung cancer, particularly in high-risk patients. The procedure involves X-ray images of the lungs taken while the patient lies on a table. The US Preventive Services Task Force recommends annual scans for patients over 50 with a 20-pack-year smoking history, but unfortunately, less than 5% of eligible individuals undergo screening in most US states.
New research suggests that many patients who are screened for lung cancer do not receive adequate long-term care, including follow-up appointments. Additionally, lung cancer diagnoses seem to be increasing among people who have never smoked or only smoked lightly.
Low-dose CT scans are an effective tool for detecting lung cancer, especially in high-risk groups such as smokers. However, they are not perfect. There is room for improvement in terms of accuracy and extendibility to other populations. For example, currently, 3-4 scans are needed to get an accurate result from a low-dose CT scan.
The Sybil model differs from traditional methods in that it only needs a single low-dose CT scan of the chest to forecast the likelihood of developing lung cancer in the next 1-6 years after screening.
The study’s researchers say that Sybil gives a risk score, not a diagnosis. This means it can be helpful in determining high-risk patients who need to be closely monitored for cancer. Recently, scholars have released the Sybil algorithm and its image annotations to the public sector in hopes of further advancing research and use in clinical settings. To support these efforts, a study was published in the Journal of Clinical Oncology.
Diagnosing low to high-risk lung cancer
A deep-learning Artificial Intelligence model was created by researchers utilizing data from 15,000 participants. The dataset consisted of 35,001 low-dose CT scans – 6,282 used to test the model and 29,719 used for training. To assist in the development of the model, two thoracic radiologists marked any suspicious lesions on scans taken from patients who later developed cancer within one year of their scan.
Through a series of tests using low-dose CT scans, the researchers found that Sybil had an allocative accuracy of 92% after one year and 86% after two. After 6 years, its probability (C-index) was 75%. Interestingly, there was no observable effect on performance due to sex, age, or smoking history.
The research team then conducted additional tests on datasets from Massachusetts General Hospital (MGH) in Boston and Chang Gung Memorial Hospital (CGMH) in Taiwan. Notably, CGMH did not require proof of smoking behavior for a low-dose CT scan; nonetheless, the predicted performance was stable across all tests.
Sybil correctly predicted 86% of lung cancer cases or healthy lungs within one year from the MGH dataset, alongside 94% of cases in the CGMH data. It also predicted 81% of lung cancers or healthy lungs among the MGH cohort and 80% among the CGHM cohort after six years. The researchers noted that Sybil was able to accurately predict traditional risk factors, such as smoking, from medical scans.
What are the areas of improvement?
The model demonstrated by the researchers is limited in certain aspects. Of the 92% of training data used, most of it came from White patients and may not be reflective of more diverse populations.
Additionally, the scans upon which Sybil’s performance was based were obtained between 2002 and 2004. Therefore, advancements in CT technology since then could potentially lead to a decreased accuracy rate. The team also noted that without detailed smoking data for CGMH patients, any conclusions about Sybil’s ability to accurately predict lung cancer among non-smokers would be speculative.
Moving forward, more research is needed to further understand the performance of Sybil in diverse populations and its clinical benefit.
Dr. Sheena Bhalla, assistant professor at Simmons Cancer Center at U.T. Southwestern Medical Center, commented that low-dose CTs have the potential for false positive results which may lead to unnecessary procedures in some patients—this should be taken into account when assessing Sybil’s efficiency and clinical impact.
The results of Sybil are indeed promising, however, there are a few considerations related to this development.
Firstly, the image quality needs to be considered to accurately assess its algorithm performance. In addition, what features within the image are most relevant to make actionable decisions? Finally, considering that there is much skepticism regarding Artificial Intelligence and Machine Learning for clinical care, it could present some translational challenges when applied.
Follow-up scans by Sybil
The researchers at Sybil believe that their software can play its part in reducing follow-up scans and biopsies among low-risk patients. Essentially, Sybil could give readings from CT scans as soon as they are available without a radiologist present – reducing the workload for doctors. This process is similar to getting a colonoscopy done and being given another only when you need it 10 years later.
The radiology screening AI tool has also been developed to identify potential abnormalities for follow-up. However, it’s important to note that a one-size approach may not be suitable for every patient and Sybil can help in personalizing screening regimens.