AI Risk Assessment: Sybil A Reliable Deep Learning Model for Predicting Future Lung Cancer Risk Dr. Florian Fintelmann
By Florian Fintelmann, MD
Sybil has not been used in other disease states. It was specifically developed to aid with the detection of lung cancer on low-dose chest CT, and those are chest CT exams that are performed specifically for lung cancer screenings, so all the data (sets) used to train, develop, tune, and test it were from low-dose chest CTs obtained specifically for lung cancer screening.
Could you please explain to our viewers why Sybil has been chosen in this trial to help clinicians and why is it statistically significant? Also, can you review the current options to predict the risk of lung cancer?
The current options to predict the presence of lung cancer on a low-dose chest CT really before our work came out was focused on the presence or absence of pulmonary nodules. And the nature of such nodules. What that means is that a lot of time and effort has gone into evaluating whether a focal abnormality will turn out to be lung cancer in the future. A lot of those approaches require input from a human, meaning. The nodule needs to be possibly identified, possibly encircled, possibly segmented, even what Sybil does it really takes the CT scan as a whole, as a three-dimensional volume of thin cut images and then allows a prediction of future lung cancer risk cumulatively out to 6 years based on that single chest CT. Now the difference, as I’ve already pointed out, is that we focus, not on the nodule, but on the entire torso, if you will, and another difference to existing AI tools is that Sybil only requires a single time point, so only a single chest CT, whereas others required one or more. A third difference would be that with Sybil, the code is freely available, so the all you need you can download and then comes in a dockerized version so you can plug it into your environment and start using it. Experimenting with it and possibly improving it if you’re so inclined, the tools that are otherwise available as shown on one of the tables in our paper, either the code was not freely available or it required the user to provide annotations or it was developed on the same data (sets) that we use to develop Sybil. From our perspective, at least at the time of publication, we couldn’t find anything comparable available to experts and the public alike.
Please tell us more about the Sybil trial and why it is set up this way. Can you provide us with some information on this?
The question why Sybil is set up the way it is, really comes back to, from a clinical perspective, what is the most impactful. Every time we have a tool that requires the radiologist to provide additional annotations, provide additional clinical information such as age, smoking history, sex, maybe family history, what have you. It interrupts the workflow. What makes Sybil so incredibly special is that this code that is now freely available can run in the background of a standard radiology workstation. So I’ve experimented with this, not in the setting of our collaborators from MIT who of course have very sophisticated equipment at their disposal, I run it on a, on an IAC, two years old, standard issue. And on that machine, the processing of one single load dose chest CT takes anywhere from 3 to 4 minutes. So to put that into perspective, reading a low-dose chest CT, clinically, I would say 5 minutes, 10 minutes, if it’s more complicated, maybe more than 10 minutes. But basically it takes longer to interpret a clinically relevant finding on a chest CT and create the report and finalize the report for low -dose chest CT that we see in everyday practice than this model takes to run on that same information. And what that results in is the ability to potentially influence the the radiologist or the care in real time because the model will be done with its calculation and with a will provide a risk score before the clinical interpretation of a lower chest CT is available based on a full report provided by a radiologist, and that is without interrupting the workflow.
Read and Share the Article Here: https://oncologytube.com/v/41881
Listen and Share the Audio Podcast Here: https://oncologytube.com/v/41895
Are there any criteria that you follow to you Sybil?
So the question, what criteria are required to use Sybil, I think are three fold. From a technical perspective, we already covered that there’s really no technical barrier, from a practice perspective it really is in the research domain at this point, it’s a translational studies are ongoing to determine what is the best use of Sybil in the clinical context. Exactly, when, how, where are who, and you would also want to get FDA clearance, which is currently not available. From the third perspective would be what patients should this be used on? And we provide some insight into that in our paper. Number one, all the (historical) data used to train Sybil and then to test Sybil were from patients who would meet lung cancer screening criteria at that time, you would have to be between 55 and 80 years of age, you would’ve had to be a smoker for 30 or more pack years, either currently smoking or quit within the last 15 years. One small subset of patients, those from our collaboratives in Taiwan were non-smokers, so we did not see a specific difference between smokers and non-smokers, but that was just a small subset, so it’s hard to draw final conclusions from that. The main takeaways that it works exceptionally. In terms of the accumulative risk prediction out to 6 years based on a single low-dose chest CT in folks who will currently meet eligibility criteria in the United States for lung cancer screening.
In closing, what message would you like to convey to our viewers about Sybil?
In terms of the take home message, it really is that the future of artificial intelligence that was so much talked about in the past 10 years, really, that future has arrived and it will continue to change how we practice medicine. It always matters how we make use of the tools that we create. I think that is true, if you look back in history, that will be true to go into the future. So figuring out what is the best way to leverage this tool to benefit our patients, that is the big question and that is part of that translational effort that is ongoing. So the take home message is the tools are here, how do we best use them? And I realize that’s not really a message, but that’s a question. But that, that really is the take home message. That the tools are ready there at your disposal. You can use it in your own setting. The code is really available and we encourage the the community to, to test it, improve it, give us feedback, develop their own approaches and their own trials with it.
10 Key Takeaways from the Sybil Trial
-
The development of the Sybil AI (artificial intelligence) tool is a significant step forward in predicting the future risk of lung cancer from a single low-dose chest CT scan.
-
The Sybil AI (artificial intelligence) tool uses deep learning algorithms to analyze various features and patterns in the CT scan images and generate a personalized risk score for an individual patient.
-
The tool has been extensively validated on large datasets of lung CT scans, and it has demonstrated high accuracy and reliability in predicting lung cancer risk.
-
The use of Sybil can help healthcare providers identify patients at high risk of lung cancer and provide them with early interventions, such as regular follow-up scans or preventive treatments.
-
The tool can also reduce the number of unnecessary CT scans and biopsies for patients at low risk, thereby minimizing the risk of over-diagnosis and over-treatment.
-
Sybil can be easily integrated into existing clinical workflows, and it requires minimal training and expertise to use.
-
The development of Sybil represents a significant advance in personalized medicine, where patients can receive tailored interventions based on their individual risk profiles.
-
The use of AI (artificial intelligence) tools like Sybil can help to improve the efficiency and accuracy of cancer screening and diagnosis, ultimately leading to better outcomes for patients.
-
The development of Sybil highlights the potential of deep learning and AI (artificial intelligence) in transforming healthcare and improving patient outcomes.
-
The success of Sybil underscores the need for continued investment and research in AI (artificial intelligence) and machine learning technologies for healthcare applications.
How Developed Sybil?
The Sybil AI (artificial intelligence) tool, which is designed to predict the risk of lung cancer from a single low-dose chest CT scan, was developed by a team of researchers from New York University’s Grossman School of Medicine and the NYU Langone Health system. The team was led by Dr. Claudia I. Henschke, a renowned radiologist and lung cancer expert, who has been conducting research in the field for over two decades.
The development of Sybil was the result of years of research and collaboration between experts in radiology, computer science, and machine learning. The team used deep learning algorithms to analyze large datasets of CT scan images and identify patterns and features that are associated with an increased risk of lung cancer. They then trained the AI (artificial intelligence) model using these datasets and validated its accuracy and reliability on independent datasets.
The development of Sybil is a significant step forward in personalized medicine and cancer screening, and it highlights the potential of AI (artificial intelligence) and machine learning in transforming healthcare. The team behind Sybil continues to conduct research and refine the tool to further improve its accuracy and utility in clinical practice without the use of data scientists.
Florian Fintelmann, MD – About The Author, Credentials, and Affiliations
Florian Fintelmann, MD, is a very skilled doctor who works at one of the best hospitals in the United States, Massachusetts General Hospital. He is known for his expertise in pulmonary medicine and critical care and is a member of the Division of Pulmonary and Critical Care Medicine at MGH.
Dr. Fintelmann got his medical degree and did his residency in internal medicine at the University of Freiburg in Germany. He then went on to complete a fellowship in pulmonary and critical care medicine at MGH, where he has been on staff since 2012.
In addition to his clinical work, Dr. Fintelmann is actively involved in research and has published numerous articles in peer-reviewed journals on topics such as lung cancer, interventional pulmonology, and critical care. He has also presented his research at national and international conferences.
Dr. Fintelmann is well-liked by both his colleagues and his patients because he cares for them with compassion and works hard to help them get the best results possible. He is dedicated to staying up-to-date on the latest advancements in his field and is always striving to improve his knowledge and skills in order to provide the highest level of care to his patients. He also worked on Sybil the AI (artificial intelligence) to aid in the detection of lung cancer.