Sybill, an artificial intelligence tool, has been developed to estimate the risk of lung cancer. Lung cancer is the world’s deadliest cancer, accounting for 1.7 million fatalities in 2020, killing more people than the following three deadliest cancers combined. Consequently, it is critical to have an early detection solution to provide immediate treatment.
Cancer early identification AI tools can result in a better long-term outcome, according to MIT’s Abdul Latif Jameel Clinic for Machine Learning in Health, Mass General Cancer Centre (MGCC), and Chang Gung Memorial Hospital (CGMH). When it’s advanced, the five-year survival rate for the lung cancer patient is closer to 70%, compared to 10% when it’s early.
“It’s the worst cancer because it’s so common and so difficult to treat, especially once it’s advanced,” explained Florian Fintelmann, MGCC Thoracic Interventional Radiologist and Co-author of the current study.
Today’s images of the lung computed tomography (LDCT) procedure is presently the most common way people are checked for lung cancer to detect it early enough to be surgically removed. But Sybill takes the screening a step further in comparison to LDCT. It can forecast the likelihood of a patient acquiring lung cancer within six years by analysing LDCT imaging data without the intervention of a radiologist.
Co-author Peter Mikhael, an MIT PhD student in electrical engineering and computer science and an affiliate of Jameel Clinic and the MIT Computer Science and Artificial Intelligence Laboratory (CSAIL), associated the procedure with “trying to identify a needle in a haystack”. However, Sybill successfully detects early-stage cancer with satisfactory results, as shown in a new article published in the Journal of Clinical Oncology.
Fintelmann and his team labelled hundreds of CT scans with evident cancerous tumours that would be used to train Sybil before testing the model on CT scans with no discernible evidence of disease. The researchers took precautionary measures to ensure Sybil’s ability to identify cancer risk appropriately.
Sybil achieved C-indices of 0.75, 0.81, and 0.80 using a heterogeneous group of lung LDCT scans gathered from the National Lung Cancer Screening Trial (NLST), Mass General Hospital (MGH), and CGMH over six years. Models with a C-index score of more than 0.7 are regarded as good and models greater than 0.8 are considered strong, with 1.00 being the maximum attainable score. The ROC-AUCs for Sybil’s one-year prediction were considerably higher, ranging from 0.86 to 0.94.
Jeremy Wohlwend, an MIT electrical engineering and computer science PhD student and Jameel Clinic and CSAIL collaborator, was shocked by Sybil’s excellent score despite the absence of apparent disease. “We discovered that even while we [as humans] couldn’t see where the cancer was, the model could still predict which lung would eventually get cancer,” he described. “It was incredibly interesting that [Sybil] could identify which side was the more likely side.”
The 3D aspect of lung CT scans made Sybil challenging to create. Because early-stage lung cancer covers minuscule areas of the lung. It is just a fraction of the hundreds of thousands of pixels that make up each CT scan. The radiology data used to train Sybil was essentially free of any indicators of malignancy. Lung nodules are denser areas of lung tissue that, while they have the potential to be malignant, are most of the time not and can be caused by healed infections or airborne irritants.
In the United State, many patients diagnosed with lung cancer today have never smoked or are former smokers who quit more than 15 years ago – characteristics that preclude both groups from receiving lung cancer CT screening in the United States. However, cancer can affect a young, healthy, and athletic individual. As a result, prevention is vital to saving more lives.
