New Study: AI Predicts the Spread of Lung Cancer to the Brain

Artificial intelligence (AI) may be useful for predicting disease progression and brain metastases in patients with lung cancer, according to a study recently published in The Journal of Pathology. This research introduces the concept that clinicians could use AI as a tool to predict disease outcomes and assist in treatment decisions in the future.

Advances in science and medicine have led to earlier detection methods and better treatment options for many types of cancer. If caught early, surgery may successfully treat or even cure the disease. However, in most cases, cancer becomes notoriously difficult to treat after it spreads (metastasizes) to other parts of the body. This includes lung cancer, which remains the leading cause of cancer deaths in the US.

The most common type of lung cancer is non-small cell lung cancer (NSCLC), which eventually spreads to the brain in nearly half of cases. Currently, clinicians do not have a reliable way to predict which patients are more likely to progress to metastatic (stage 4) disease. This presents a tough dilemma for patients with early-stage disease after having lung surgery to remove the tumor: Is it better to take a watchful-waiting approach and risk swift recurrence — or use a proactive approach involving systemic treatments like chemotherapy and risk unnecessary side effects?

Now, an AI algorithm shows promise as a tool that may help predict the spread of lunger cancer and therefore guide the decision-making process. The new study was conducted by researchers from Caltech and Washington University School of Medicine in St. Louis.

“Our study focuses on the predominant type of lung cancer. Most of these [NSCLC cases] are caused by smoking, but about 30% of cases occur in non-smokers,” Richard J. Cote, M.D., principal investigator and senior author of the study, told Managed Healthcare Executive in an interview on March 11. Cote is an Edward Mallinckrodt Professor and serves as the chair of the department of pathology and immunology at WashU Medicine.

Cote emphasized their study’s focus on patients with stage 1 NSCLC, an early stage of the disease that is potentially curable via surgery or a combination of surgery and chemotherapy or other treatments. “Stage one is where the greatest management dilemma exists,” he explained.

Traditionally, pathologists read microscope-derived images of biopsied lung tissue to look for markers of cancer, such as atypical cell characteristics. The new study sought to determine if a deep-learning AI network could be used to aid pathologists in reading the histologic images as well as explore the potential of AI to detect specific features or abnormalities in the biopsied tissue images that could predict metastatic progression.

“Our question was: When a patient first comes in, at the initial diagnostic biopsy before treatment, could we train AI and use AI to understand whether or not that patient is destined to develop brain metastases, or whether that patient is destined to not develop metastases at all?” Cote explained.

To address this question, the researchers utilized biopsy data from 158 patients with stage 1-3 NSCLC who were monitored for brain metastases for at least 5 years. Of these, 65 patients had brain metastases (Met+) and 93 did not (Met-). Diagnostic slides were digitized for analysis. The AI algorithm was trained and validated using 118 cases (45 Met+, 73 Met-) and tested on 40 separate cases (20 Met+, 20 Met-).

The results revealed that AI was able to correctly predict the development of brain metastases with 87% accuracy from reading the biopsied samples. In comparison, four trained pathologists correctly predicted cancer progression about 57% of the time when given the same task.

Cote highlighted two key points from the results. First, the algorithm excelled in reliably distinguishing patients who developed brain metastases versus those patients who never developed any metastases. Second, the algorithm was even more reliable in predicting those patients with stage 1 disease who did not develop any metastases.

This technology may “provide a clear management pathway for a relatively large subset of patients with non-small cell lung cancer,” Cote said. “In my view, this is the most actionable point—potentially sparing patients with early-stage NSCLC the expense and toxicity of further unnecessary treatment.”

While the results of previous studies already support a role for AI in diagnosis, this is the first study to go a step further and ask AI to predict disease progression in early-stage NSCLC.

While promising, more research is necessary before AI predictions can be implemented in clinical practice. Larger studies involving data from multiple institutions are needed to validate and further improve the predictive abilities of the algorithm.

“We have quite a bit more work to do. But this is a really exciting first step,” Cote said.

Cote and co-senior author Changhuei Yang, Ph.D., of WashU, conducted this research alongside a a team of pathologists, scientists, and engineers from WashU and Caltech. The study was funded by the Heritage Medical Research Institute, the Caltech Center for Sensing to Intelligence, WashU’s Personalized Medicine Initiative, and the National Cancer Institute.