Using Lung Scans to Predict Treatment Response

By Sara Bondell - June 18, 2020

Immunotherapy is one of the most successful treatment options for advanced non-small cell lung cancer. However, only about half of patients respond to it and many suffer from adverse side effects.

“It is very important to determine biomarkers at the beginning of treatment to predict whether a patient will respond to it,” said Dr. Robert Gillies, a researcher in the Departments of Cancer Physiology and Radiology at Moffitt Cancer Center. “You don’t want to treat patients with a therapy they won’t respond to because all therapies have associated toxicities.”

Dr. Robert Gillies, Departments of Cancer Physiology and Radiology
Dr. Robert Gillies, Departments of Cancer Physiology and Radiology
Dr. Matthew Schabath, Departments of Cancer Epidemiology and Thoracic Oncology
Dr. Matthew Schabath, Departments of Cancer Epidemiology and Thoracic Oncology

Gillies, along with co-leader Dr. Matthew Schabath, a researcher in the Departments of Cancer Epidemiology and Thoracic Oncology, conducted two studies to show that analysis of clinical images obtained prior to treatment could be used to predict clinical benefit among lung cancer patients receiving immunotherapy. These studies, presented at the annual American Association for Cancer Research meeting, show that radiomics—quantitative image analysis that extract features from computed tomography (CT) scans or positron emission tomography (PET) scans—can be used to improve treatment decisions in cancer.

The first study analyzed scans and clinical data of more than 300 patients who received single or double agent immunotherapy. The researchers used an image-based model to assign the patient into four risk groups based on survival. Then, they looked at specific gene-expressions that could be related to their image-based signature to reveal the underlying biology.

“We found that those in very high-risk group should not be treated with immunotherapy because they will not respond,” said Moffitt postdoctoral fellow Dr. Ilke Tunali. “We also found that our image-based biomarker was associated with specific gene-expression, related to tumor hypoxia, and was associated with poor outcomes in lung cancer patients not treated with immunotherapy.”

Tunali says further study is needed but the image-based biomarker finding could have broader clinical benefit and help determine if other therapies, not just immunotherapy, will work for a patient.

The second study investigated if a deep learning software, developed by Moffitt postdoctoral fellow Dr. Wei Mu, could use PET and CT scans to predict PD-L1 expression as well as response to immunotherapy. PD-L1 protein expression in cancer cells is an accepted biomarker to predict a positive response to immunotherapy. Currently, the only way to determine PD-L1 expression is with a biopsy, which itself has some risk for lung cancer patients.

This study involved 546 patients whose images were retrospectively analyzed with a deep learning Convolutional Neural Net (CNN) in training and independent validation cohorts. A computer algorithm was developed that could predict PD-L1 expression with high accuracy.  When this algorithm was applied to a separate cohort of patients, it was able to also accurately predict whether a patient would respond to immunotherapy, offering an easier and less invasive prediction option. 

The Moffitt researchers are now taking the decision support tool one step further to determine which treatment type is best for a patient. Tyrosine kinase inhibitors (TKIs) are a targeted therapy for lung cancer patients with the most common mutation, in the EGF Receptor. Patients who respond well to TKIs usually don’t respond well to immunotherapy and vice-versa, and some patients do not respond to either therapy. Using the similar deep learning program and standard-of-care imaging, the researchers have now been able to predict if a patient has an EGFR mutation and what their PD-L1 score is. The results then indicate which treatment to use.

“This is exciting because it is truly the first time that we can use radiomics for a decision support tool to decide between two types of treatments,” said Gillies. “Instead of just predicting if you are going to do well, we can say which therapy is best.”

Contact the Author

Sara Bondell Medical Science Writer More Articles


Most Popular