Lung cancer is a devastating disease. According to the World Health Organizationlung cancer is one of the most common causes of death worldwide, accounting for nearly 2.21 million cases in 2020 alone. Importantly, the disease can be progressive; that is, for many, it may start out as just mild symptoms that raise no alarm, before quickly evolving into a life-threatening diagnosis, leading to death. Fortunately, the range of therapeutics focused on helping patients with lung cancer has grown tremendously in the last two decades. However, early detection of the cancer is still one of the only means to significantly decrease mortality rates.
One notable accomplishment in this arena is the recent announcement by the Massachusetts Institute of Technology (MIT) and Mass General Hospital (MGH) regarding the development of a deep learning model named “Sybil” that can be used to predict lung cancer risk, using data from just a single CT scan. The study was formally published in the Journal of Clinical Oncology last week, and discusses how “tools that provide personalized future cancer risk assessment could focus approaches toward those most likely to benefit.” Hence, the study leaders posited that “a deep learning model assessing the entire volumetric LDCT [Low Dose Contrast CT] data could be built to predict individual risk without requiring additional demographic or clinical data.”
The model starts with a basic tenet: “LDCT images contain information that is predictive of future lung cancer risk beyond currently identifiable features such as lung nodules.” Hence, the developers sought to “develop and validate a deep learning algorithm that predicts future lung cancer risk out to 6 years from a single LDCT scan, and assess its potential clinical impact.”
Overall, the study has been remarkably successful, thus far: Sybil is able to predict a patient’s future lung cancer risk to a certain extent of accuracy, using the data from just one LDCT.
Without a doubt, clinical applications and implications for this technology are still immature. Even the study leaders agree that significant work will need to be done to figure out exactly how to apply this technology in actual clinical practice— specifically with regards to developing a degree of confidence in the technology, with which physicians and patients will feel safe relying on the system’s outputs.
However, the premise of the algorithm is still incredibly powerful and entails a potential game-changer in the realm of predictive diagnostics.
Diagnostic measures have never before been so powerful. The fact that a tool can use just one CT scan to predict a long-term disease function could potentially solve many problems— the most important of which is enabling early treatment and decreased mortality.
Pundits, at initial blush, may push back against systems like these, remarking that no AI system could possibly match the judgement and clinical prowess well enough to replace a human physician. But the purpose of systems like these is not necessarily to replace physician expertise, but rather to potentially augment physican workflows.
A system like Sybil could very easily be used as a recommendation tool, flagging potentially concerning CTs to a physician, who could then use their own clinical judgement to either agree or disagree with Sybil’s recommendation. This would not only likely improve clinical throughput, but could also act as a secondary “check” process and possibly enhance diagnostic accuracy.
Undoubtedly, there is still a lot of work to be done in this arena. Scientists, developers, and innovators have a long journey ahead of them in not only perfecting the actual algorithm and system itself, but also in navigating the hyper-nuanced arena of introducing this technology into actual clinical applications. Nevertheless, the technology, the intention, and the potential it holds with regards to bettering patient care, if it is developed in a safe, ethical, and efficacious manner, is indeed promising for the generation of diagnostics to come.