Meet Sybil: The AI That Sees Lung Cancer Before It Strikes

MIT’s Jameel Clinic has developed Sybil, an AI-powered tool that analyzes CT scans to predict a patient’s risk of developing lung cancer up to six years in advance. Unlike traditional methods that rely on age and smoking history, Sybil detects subtle patterns across the entire scan—beyond what human eyes can see. Tested on thousands of scans, Sybil identified nearly 90% of future lung cancer cases by screening just the top 20% at-risk patients. This breakthrough could personalize screening, catch cancer earlier, and reduce invasive biopsies, potentially transforming lung cancer detection worldwide.

Transcript

00:00:00 (suspenseful music) - [Narrator] Lung cancer is the leading cause of cancer death in both men and women worldwide. One of the major challenges in thinking about lung cancer screening, is that we don't have a good tool to tell us who is at risk for the disease. - The risk calculators for lung cancer risk that exist are heavily based on things like tobacco exposure, and we know that the demographics

00:00:29 of lung cancer is changing. And so, if we only focus on putting into a calculator, how old are you, how many years did you smoke tobacco? We know we're gonna come up with faulty information. - [Narrator] At MIT's Jameel Clinic, PhD student Peter Mikhael and Professor Regina Barzilay, along with researchers at MGH have developed an AI called Sybil that has become an innovative new tool for doctors.

00:00:56 - Sybil can do what human eyes cannot detect, to look at the very subtle changes in the images and predict the likelihood of the patient developing disease in the future. - [Narrator] Cross-sectional images from CT scans are traditionally used to tell us what is going on with a patient in the present. Is there an infection, a broken bone, is cancer visible? But the CT images hold a treasure trove of additional information.

00:01:23 Subtle nodules, architecture of the lung, even signs from other parts of the body might signal an increased likelihood of future cancer development. (lively upbeat music) - We really don't have a precedent for getting a scan today and using it to say, "Well, three years from now is this patient gonna have cancer?" Sybil, we think is a tool that can help illuminate for doctors

00:01:51 who are those people who are at risk. - Sybil can look at a CT scan of a patient and tell us the likelihood of this patient developing disease in one, two, and even up to six years. - [Narrator] Instead of taking into account clinical factors like age and tobacco exposure, Sybil analyzes all of the information available in a patient's CT scan. By drawing on knowledge from thousands of other CT scans that were used to train the model,

00:02:20 Sybil can classify the image as a high-risk person or a non high-risk person. Then, the doctor is able to make a clinical decision. - The power of Sybil really lies in integrating all that information and not just the information that we as radiologists have sort of defined as relevant over time. - So, this is an example where Sybil really was able to tell the future in a way. This is the same guy two years apart,

00:02:54 and here, the radiologist didn't see a lung cancer. - [Radiologist] As a radiologist, I wouldn't have pointed to this at all. - But Sybil was able to say, something about this scan makes me worried, especially this little area. - In that location. Correct. - And then the same guy, two years later, there's actually, a lung cancer in that area. - Yeah.

00:03:16 - [Narrator] Sybil's power is validated across three independent datasets. Here's a closer look at how the model performed when tested on data from the National Lung Screening Trial. Out of 2,127 participants that had conventional screening, 39 of them were diagnosed within a one year period after their initial screen. Sybil was asked to identify the top 20% of participants with the highest lung cancer risk. That group included 35 of the 39 cases

00:03:46 from the National Lung Screening Trial. By honing in on the top 20% of risk, Sybil was able to identify nearly 90% of the cancers a year prior to the actual diagnosis and by screening only a fifth of the cases used in conventional methods. The ability to more efficiently diagnose lung cancer risk earlier can help make eligibility criteria for lung cancer screening less restrictive. - When I talk to patients or other providers

00:04:16 and tell them about Sybil, the first question they always ask is, "Well, what is the machine looking at?" But as medical providers, we're trying to wrap our head around this whole new construct of AI where the data is so vast and we're putting in the entire volumetric data in a CAT scan, it's not just the lungs, it's the heart, it's the muscles, it's the bones, it's other organs, everything.

00:04:44 You know, the skin, the subcutaneous fat, everything that's in our body is being examined by Sybil. And the way that machine learning works is you don't tell it. You don't tell it, I think it's most important to look at the lungs or I think it's most important to look at, you know, the bones. The model actually, looks at thousands and thousands of scans. We do tell it who has cancer and who doesn't,

00:05:08 and it picks out on its own what are the features that it's gonna look at. And probably not in the same way that a human radiologist categorizes things into bones and heart and lungs. - One unique advantage of AI tools like Sybil is the ability to look and keep in its memory, you know, hundreds of thousands of images, way more than what any human radiologist would see in their lifetime.

00:05:32 And not only remember them, but also know their outcomes. Being able to abstract these images to generalize to new patients is a unique power of these type of tools to solve very hard prediction tasks. - So, it is a really different paradigm. It's a completely new idea for how to use a radiology scan. - [Narrator] Researchers are exploring Sybil's potential clinical applications, which hold promise for more personalized cancer screenings

00:06:05 that could reduce the need for invasive biopsies. These improvements may allow doctors to personalize screening efforts, focusing on the patient's most at risk. Further research involves elevating equitability by training Sybil on more diverse datasets. Ensuring it serves all patient populations effectively. - We believe that we will change the future of lung cancer screening, and already have. We know that Sybil is already being picked up

00:06:32 and implemented in hospitals in multiple countries. We're working on future studies and really integrating Sybil into the clinical care pathway. And the results of these studies have the potential to change care immediately. - One of the amazing innovations that Regina has brought to the medical field is this idea that radiology can be used to predict future risk. I mean, that's just a novel idea, and so simple when you, you know,

00:07:05 step back and think about it, it's awesome. (gentle upbeat music)